{ "schemaVersion": "1", "generatedAt": "2026-05-04T00:00:00Z", "version": "2.0.0", "models": [ { "id": "nequip", "type": "node", "category": "Equivariant", "label": "NequIP", "year": 2021, "author": "Harvard (Kozinsky) / MIT (Smidt)", "x": 100, "y": 150, "desc": "E(3)-equivariant message-passing potential that set the template for data-efficient, high-accuracy force fields across molecules and materials.", "githubUrl": "https://github.com/mir-group/nequip", "paperUrl": "https://arxiv.org/abs/2101.03164", "coverage": [ "organic molecules", "small-molecule reactions", "materials" ], "useCases": [ "data-efficient fitting", "ab-initio MD surrogate" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE", "LAMMPS" ], "license": "MIT", "maintenance": "maintained", "lastReviewed": "2026-05-05", "trainingData": [ "custom DFT" ], "tags": [ "equivariant", "message-passing", "E(3)" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "dataset-dependent (general molecules and materials)", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "allegro", "type": "node", "category": "Equivariant", "label": "Allegro", "year": 2023, "author": "Harvard (Kozinsky lab)", "x": 100, "y": 320, "desc": "Strictly local equivariant architecture designed for massive parallel MD (100M+ atoms) while retaining NequIP-level accuracy.", "githubUrl": "https://github.com/mir-group/allegro", "paperUrl": "https://www.nature.com/articles/s41467-023-36329-y", "coverage": [ "general materials", "organic molecules" ], "useCases": [ "massive parallel MD", "large-scale MD" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE", "LAMMPS", "PyTorch" ], "license": "MIT", "maintenance": "maintained", "lastReviewed": "2026-05-05", "trainingData": [ "custom DFT" ], "tags": [ "equivariant", "strictly local", "parallel MD" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "dataset-dependent (general molecules and materials)", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "eqv2", "type": "node", "category": "Transformer", "label": "Equiformer V2", "year": 2024, "author": "Meta FAIR / MIT (Liao, Smidt et al.)", "x": 380, "y": 150, "desc": "Improved equivariant transformer with higher-degree tensor representations; achieves state-of-the-art OC20/OC22 performance with strong data-efficiency.", "githubUrl": "https://github.com/atomicarchitects/equiformer_v2", "paperUrl": "https://arxiv.org/abs/2306.12059", "coverage": [ "catalysts", "surfaces", "general materials" ], "useCases": [ "catalyst discovery", "surface chemistry" ], "properties": [ "energy", "forces" ], "frameworks": [ "PyTorch", "ASE" ], "license": "MIT", "maintenance": "maintained", "lastReviewed": "2026-05-05", "trainingData": [ "OC20", "OC22" ], "tags": [ "transformer", "equivariant", "attention", "higher-order tensors" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements covered by OC20 / OC22 (~56 elements)", "equivariance": "constrained", "architecture": "gnn", "usesAttention": true, "longRange": false, "numElements": 56 }, { "id": "mace", "type": "node", "category": "Equivariant", "label": "MACE", "year": 2022, "author": "Cambridge (Csányi) / UBC (Ortner) — Batatia et al.", "x": 660, "y": 150, "desc": "Higher-order equivariant message passing (4-body messages) that reaches SOTA accuracy with only 1-2 layers; later extended into the universal MACE-MP foundation model family.", "githubUrl": "https://github.com/ACEsuit/mace", "paperUrl": "https://arxiv.org/abs/2206.07697", "coverage": [ "general materials", "organic molecules", "oxides" ], "useCases": [ "universal MLIP", "MD at scale", "high-throughput screening" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE", "LAMMPS" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-05", "trainingData": [ "MPTrj", "Alexandria" ], "tags": [ "equivariant", "higher-order", "foundation model" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements covered by MPTrj / Alexandria (~89 elements)", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "grace", "type": "node", "category": "Equivariant", "label": "GRACE", "year": 2024, "author": "ICAMS", "x": 660, "y": 320, "desc": "Graph Atomic Cluster Expansion: a foundation-scale implementation of ACE with explicit multi-body basis functions for wide-coverage materials modelling.", "githubUrl": "https://github.com/ICAMS/grace-tensorpotential", "paperUrl": "https://arxiv.org/abs/2508.17936", "coverage": [ "general materials" ], "useCases": [ "universal MLIP", "high-throughput screening" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE", "LAMMPS" ], "license": "Apache-2.0", "maintenance": "active", "lastReviewed": "2026-05-05", "trainingData": [ "OMat24", "sAlex", "MPTrj" ], "tags": [ "equivariant", "ACE", "graph", "foundation model" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements covered by OMat24 / MPTrj (~89 elements)", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false, "numElements": 89 }, { "id": "orb", "type": "node", "category": "Transformer", "label": "Orb-v3", "year": 2025, "author": "Orbital Materials", "x": 950, "y": 150, "desc": "Non-equivariant, non-conservative graph neural network potential systematically trading off roto-equivariance, conservatism, and graph sparsity for >10x latency and >8x memory reduction at near-SOTA accuracy on large periodic systems.", "githubUrl": "https://github.com/orbital-materials/orb-models", "paperUrl": "https://arxiv.org/abs/2504.06231", "coverage": [ "general materials", "periodic crystals" ], "useCases": [ "large-cell MD", "high-throughput screening" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE" ], "license": "Apache-2.0", "maintenance": "active", "lastReviewed": "2026-05-05", "trainingData": [ "MPTrj", "Alexandria" ], "tags": [ "transformer", "throughput", "torch.compile" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements covered by MPTrj / Alexandria (~89 elements)", "equivariance": "learnt", "architecture": "gnn", "usesAttention": true, "longRange": false, "numElements": 89 }, { "id": "orb_v2", "type": "node", "category": "Transformer", "label": "Orb-v2", "year": 2024, "author": "Orbital Materials", "x": 3740, "y": 150, "desc": "Predecessor to Orb-v3. A non-equivariant graph network exploring trade-offs between accuracy and inference cost on materials simulation; trained on MPTrj + Alexandria with rotation-invariance learnt rather than imposed.", "githubUrl": "https://github.com/orbital-materials/orb-models", "paperUrl": "https://arxiv.org/abs/2410.22570", "coverage": [ "materials" ], "useCases": [ "large-cell MD", "high-throughput screening" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE", "PyTorch" ], "license": "Apache-2.0", "maintenance": "maintained", "lastReviewed": "2026-05-06", "trainingData": [ "MPTrj", "Alexandria" ], "tags": [ "learnt-equivariance", "graph", "foundation" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements covered by MPTrj / Alexandria (~89 elements)", "equivariance": "learnt", "architecture": "gnn", "usesAttention": true, "longRange": false, "numElements": 89 }, { "id": "orbmol", "type": "node", "category": "Transformer", "label": "OrbMol", "year": 2025, "author": "Orbital Materials", "x": 950, "y": 320, "desc": "Orb-v3 variant for molecules, electrolytes, metal complexes, and biomolecules, trained on the ~100M-structure OMol25 dataset with explicit total-charge and spin-multiplicity conditioning.", "githubUrl": "https://huggingface.co/Orbital-Materials/OrbMol", "paperUrl": "https://www.orbitalmaterials.com/posts/orbmol-extending-orb-to-molecular-systems", "coverage": [ "organic molecules", "electrolytes", "metal complexes", "biomolecules" ], "useCases": [ "drug discovery", "electrolyte design", "molecular MD" ], "properties": [ "energy", "forces" ], "frameworks": [ "ASE" ], "license": "Apache-2.0", "maintenance": "active", "lastReviewed": "2026-05-05", "trainingData": [ "OMol25" ], "tags": [ "transformer", "charge-aware", "spin-aware", "foundation model" ], "supportsCharges": true, "supportsSpins": true, "elementsCovered": "elements present in OMol25 (organic + electrolyte + metal-complex chemistry)", "equivariance": "learnt", "architecture": "gnn", "trainingSetSize": 100000000, "usesAttention": true, "longRange": false }, { "id": "tfn", "type": "node", "category": "Equivariant", "label": "Tensor Field Networks", "year": 2018, "author": "Thomas et al.", "x": 380, "y": 320, "desc": "First E(3)-equivariant convolutional architecture for point clouds, introducing spherical-harmonic tensor products that underlie nearly all modern equivariant MLIPs (NequIP, MACE, Equiformer).", "githubUrl": "https://github.com/tensorfieldnetworks/tensorfieldnetworks", "paperUrl": "https://arxiv.org/abs/1802.08219", "coverage": [ "point clouds", "small molecules" ], "useCases": [ "equivariance research baseline" ], "properties": [ "energy", "forces" ], "frameworks": [ "PyTorch" ], "license": "MIT", "maintenance": "archived", "lastReviewed": "2026-05-05", "trainingData": [ "toy/synthetic" ], "tags": [ "equivariant", "E(3)", "tensor field", "point cloud" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "—", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "se3t", "type": "node", "category": "Transformer", "label": "SE(3)-Transformer", "year": 2020, "author": "Fuchs et al.", "x": 1510, "y": 150, "desc": "Self-attention generalized to SE(3)-equivariant inputs via tensor field attention; an early blueprint for equivariant transformer architectures like Equiformer.", "githubUrl": "https://github.com/FabianFuchsML/se3-transformer-public", "paperUrl": "https://arxiv.org/abs/2006.10503", "coverage": [ "point clouds", "small molecules" ], "useCases": [ "equivariant attention research baseline" ], "properties": [ "energy", "forces" ], "frameworks": [ "PyTorch" ], "license": "MIT", "maintenance": "archived", "lastReviewed": "2026-05-05", "trainingData": [ "QM9", "MD17" ], "tags": [ "transformer", "SE(3)", "self-attention", "equivariant" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "—", "equivariance": "constrained", "architecture": "gnn", "usesAttention": true, "longRange": false }, { "id": "jmp", "type": "node", "category": "Transformer", "label": "JMP", "year": 2024, "author": "Shoghi et al. (CMU / Meta)", "x": 1510, "y": 320, "desc": "Joint Multi-domain Pre-training: a strategy that trains one shared GemNet-OC backbone simultaneously on OC20, OC22, ANI-1x and Transition-1x (~120M systems), demonstrating multi-dataset pretraining for transferable potentials — a precursor to the universal MLIP foundation models that followed.", "githubUrl": "https://github.com/facebookresearch/JMP", "paperUrl": "https://arxiv.org/abs/2310.16802", "coverage": [ "catalysts", "surfaces", "organic molecules", "transition states" ], "useCases": [ "multi-task pretraining", "transferable potential fine-tuning" ], "properties": [ "energy", "forces" ], "frameworks": [ "PyTorch", "ASE" ], "license": "CC-BY-NC-4.0", "maintenance": "archived", "lastReviewed": "2026-05-05", "trainingData": [ "OC20", "OC22", "ANI-1x", "Transition-1x" ], "tags": [ "transformer", "multi-task pretraining", "foundation model" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "elements present in OC20 / OC22 / ANI-1x / Transition-1x (organic + catalytic chemistry)", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false, "trainingSetSize": 120000000 }, { "id": "esen", "type": "node", "category": "Equivariant", "label": "eSEN", "year": 2025, "author": "Meta FAIR", "x": 1230, "y": 150, "desc": "Equivariant Smooth Energy Network: conservative-force equivariant GNN with a smooth potential energy surface designed for stable long-horizon MD. Serves as the backbone underneath Meta's UMA foundation model.", "githubUrl": "https://github.com/facebookresearch/fairchem", "paperUrl": "https://arxiv.org/abs/2502.12147", "coverage": [ "general materials", "catalysts" ], "useCases": [ "long-horizon MD", "stable PES sampling" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE", "LAMMPS", "PyTorch" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-05", "trainingData": [ "OMat24", "MPTrj" ], "tags": [ "equivariant", "smooth PES", "foundation model" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements covered by OMat24 / MPTrj (~89 elements)", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false, "numElements": 89 }, { "id": "uma", "type": "node", "category": "Transformer", "label": "UMA", "year": 2025, "author": "Meta FAIR", "x": 1230, "y": 320, "desc": "Universal Model for Atoms: a Mixture of Linear Experts (MoLE) foundation model built on the eSEN backbone, trained on ~500M structures spanning OC20, ODAC23, OMat24, OMC25, and OMol25. NeurIPS 2025 spotlight.", "githubUrl": "https://huggingface.co/facebook/UMA", "paperUrl": "https://arxiv.org/abs/2506.23971", "coverage": [ "general materials", "catalysts", "MOFs", "molecular crystals", "organic molecules" ], "useCases": [ "universal MLIP", "multi-task simulation", "drug + materials" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE", "LAMMPS", "PyTorch" ], "license": "proprietary", "maintenance": "active", "lastReviewed": "2026-05-05", "trainingData": [ "OC20", "ODAC23", "OMat24", "OMC25", "OMol25" ], "tags": [ "transformer", "mixture-of-experts", "foundation model", "charge-aware", "spin-aware" ], "supportsCharges": true, "supportsSpins": true, "elementsCovered": "elements present in OC20 / ODAC23 / OMat24 / OMC25 / OMol25 (~89 elements)", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false, "numElements": 89, "trainingSetSize": 500000000 }, { "id": "nequix", "type": "node", "category": "Equivariant", "label": "Nequix", "year": 2025, "author": "Koker & Smidt (MIT Atomic Architects)", "x": 1790, "y": 150, "desc": "Compact E(3)-equivariant foundation potential pairing a simplified NequIP design with equivariant RMS layer normalization and the Muon optimizer; 700K parameters trained in ~100 A100 GPU-hours, with ~20x lower training cost and two orders of magnitude faster inference than the top Matbench-Discovery models.", "githubUrl": "https://github.com/atomicarchitects/nequix", "paperUrl": "https://arxiv.org/abs/2508.16067", "isNew": true, "coverage": [ "general materials" ], "useCases": [ "budget-constrained training", "phonon calculations" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE", "JAX-MD", "PyTorch" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-05", "trainingData": [ "MPTrj", "OMat24", "sAlex", "MDR Phonon" ], "tags": [ "equivariant", "JAX", "foundation model", "Muon optimizer" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements covered by MPTrj / OMat24 (~89 elements)", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false, "numElements": 89 }, { "id": "pet", "type": "node", "category": "Transformer", "label": "PET", "year": 2023, "author": "Pozdnyakov & Ceriotti (EPFL)", "x": 3740, "y": 320, "desc": "Point Edge Transformer: an unconstrained graph transformer for atomistic systems. Drops strict equivariance in favour of attention-based message passing on point-and-edge inputs, with rotation-invariance learnt from data rather than imposed by the architecture. Direct precursor to PET-MAD.", "githubUrl": "https://github.com/lab-cosmo/pet", "paperUrl": "https://arxiv.org/abs/2305.19302", "coverage": [ "molecules", "materials" ], "useCases": [ "graph-transformer MLIP baseline", "learnt-equivariance research" ], "properties": [ "energy", "forces" ], "frameworks": [ "PyTorch" ], "license": "MIT", "maintenance": "maintained", "lastReviewed": "2026-05-06", "trainingData": [ "custom" ], "tags": [ "transformer", "attention", "learnt-equivariance" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "—", "equivariance": "learnt", "architecture": "gnn", "usesAttention": true, "longRange": false }, { "id": "petmad", "type": "node", "category": "Transformer", "label": "PET-MAD", "year": 2025, "author": "EPFL (Ceriotti lab)", "x": 1790, "y": 320, "desc": "Lightweight universal transformer-GNN potential (Point-Edge Transformer) trained on the Massive Atomistic Diversity (MAD) dataset spanning solids, surfaces, and molecules; competitive with larger uMLIPs across diverse atomistic systems.", "githubUrl": "https://github.com/lab-cosmo/pet-mad", "paperUrl": "https://arxiv.org/abs/2503.14118", "isNew": true, "coverage": [ "general materials", "surfaces", "molecules" ], "useCases": [ "lightweight universal MLIP", "MD with uncertainty" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE", "LAMMPS" ], "license": "BSD-3-Clause", "maintenance": "active", "lastReviewed": "2026-05-05", "trainingData": [ "MAD" ], "tags": [ "transformer", "PET", "foundation model", "lightweight" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "85 elements", "architecture": "gnn", "usesAttention": true, "numElements": 85, "equivariance": "learnt", "longRange": false }, { "id": "eqv3", "type": "node", "category": "Transformer", "label": "Equiformer V3", "year": 2026, "author": "Liao et al. (MIT Atomic Architects)", "x": 2060, "y": 320, "desc": "Third-generation SE(3)-equivariant graph attention transformer with improved efficiency, expressivity, and generality; achieves the strongest results within the Equiformer family on OC20 S2EF-2M, MPtrj, OMat24, sAlex, and Matbench-Discovery.", "githubUrl": "https://github.com/atomicarchitects/equiformer_v3", "paperUrl": "https://arxiv.org/abs/2604.09130", "isNew": true, "coverage": [ "general materials", "catalysts" ], "useCases": [ "foundation MLIP", "Matbench-Discovery" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "PyTorch", "ASE" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-05", "trainingData": [ "OC20", "MPTrj", "OMat24", "sAlex", "Matbench-Discovery" ], "tags": [ "transformer", "SE(3)-equivariant", "foundation model" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements covered by MPtrj / OMat24 (~89 elements)", "equivariance": "constrained", "architecture": "gnn", "usesAttention": true, "longRange": false, "numElements": 89 }, { "id": "mace_polar1", "type": "node", "category": "Equivariant", "label": "MACE-POLAR-1", "year": 2026, "author": "ACEsuit (Kovacs, Batatia, Csanyi et al.)", "x": 2060, "y": 150, "desc": "Polarisable electrostatic foundation model that augments MACE with a non-self-consistent polarisable field formalism, learning atomic charge and spin densities (Gaussian-type multipoles) directly from energies/forces; global charge/spin constraints are enforced via learnable Fukui equilibration functions. Trained on OMol25 (~100M structures at ωB97M-V), released in M (12 A) and L (18 A) receptive-field variants for molecular chemistry and non-covalent interactions.", "githubUrl": "https://github.com/ACEsuit/mace-foundations", "paperUrl": "https://arxiv.org/abs/2602.19411", "isNew": true, "coverage": [ "organic molecules", "non-covalent interactions" ], "useCases": [ "polarisable molecular MD", "drug discovery" ], "properties": [ "energy", "forces" ], "frameworks": [ "ASE", "LAMMPS", "PyTorch" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-05", "trainingData": [ "OMol25" ], "tags": [ "equivariant", "MACE", "polarisable", "charge-aware", "spin-aware", "long-range-electrostatics", "foundation model" ], "supportsCharges": true, "supportsSpins": true, "elementsCovered": "elements present in OMol25 (molecular chemistry / non-covalent interactions)", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": true, "trainingSetSize": 100000000 }, { "id": "mace_osaka26", "type": "node", "category": "Equivariant", "label": "MACE-Osaka26", "year": 2026, "author": "Kuroda, Ishihara, Shiota, Mizukami (Osaka Univ. / QIQB)", "x": 2340, "y": 150, "desc": "Multi-domain universal MACE-architecture potential extending the MACE-Osaka series to 97 elements — the broadest elemental coverage to date — by integrating MACE-Osaka24's inorganic + organic data with the newly constructed HE26 heavy-element dataset of minor actinides assembled from experimental and computational literature. Targets nuclear and actinide chemistry while retaining strong performance on the inorganic MPtrj and organic OFF23 test sets.", "githubUrl": "https://github.com/qiqb-osaka/mace-osaka26", "paperUrl": "https://arxiv.org/abs/2603.03223", "isNew": true, "coverage": [ "nuclear materials", "actinides", "general materials", "organic molecules" ], "useCases": [ "actinide chemistry", "nuclear waste materials" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE", "LAMMPS" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-05", "trainingData": [ "MACE-Osaka24", "HE26", "MPtrj", "OFF23" ], "tags": [ "MACE", "foundation model", "actinides", "nuclear", "97 elements" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "97 elements (incl. minor actinides)", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false, "numElements": 97 }, { "id": "mlanet", "type": "node", "category": "Equivariant", "label": "MLANet", "year": 2026, "author": "Bi, Zhao, Sun, Hu, Lu, Cheng (Shanghai University)", "x": 2340, "y": 320, "desc": "Efficient equivariant graph neural network MLIP that introduces a geometry-aware dual-path dynamic attention mechanism inside its message-passing layers and a physics-informed multi-perspective pooling strategy for global system representations. Demonstrates competitive accuracy with mainstream equivariant models at markedly lower computational cost across organic molecules (QM7, MD17), Li-containing crystals, two-dimensional materials (bilayer graphene, black phosphorus), surface catalytic reactions (formate decomposition), and charged systems, while remaining stable for long-time MD simulations.", "paperUrl": "https://arxiv.org/abs/2603.22810", "isNew": true, "coverage": [ "organic molecules", "Li-containing crystals", "2D materials", "catalytic surfaces", "charged systems" ], "useCases": [ "efficient equivariant MD", "long-time MD" ], "properties": [ "energy", "forces" ], "frameworks": [ "PyTorch" ], "maintenance": "experimental", "lastReviewed": "2026-05-05", "trainingData": [ "QM7", "MD17" ], "tags": [ "equivariant", "dynamic attention", "efficient" ], "supportsCharges": true, "supportsSpins": false, "elementsCovered": "—", "equivariance": "constrained", "architecture": "gnn", "usesAttention": true, "longRange": false }, { "id": "equiewald", "type": "node", "category": "Equivariant", "label": "EquiEwald", "year": 2026, "author": "Zhang et al. (Shanghai AI Lab / CUHK)", "x": 2620, "y": 150, "desc": "Unified neural interatomic potential that embeds an Ewald-inspired reciprocal-space formulation inside an irreducible SO(3)-equivariant framework. Performs equivariant message passing in reciprocal space via learned equivariant k-space filters and an equivariant inverse transform, capturing anisotropic tensorial long-range correlations without sacrificing physical consistency; consistently improves energy and force accuracy, data efficiency, and long-range extrapolation across periodic systems, supramolecular assemblies, conjugated molecules, charged dimers, and biomolecular dynamics.", "paperUrl": "https://arxiv.org/abs/2603.18389", "isNew": true, "coverage": [ "periodic systems", "supramolecular assemblies", "biomolecules" ], "useCases": [ "long-range electrostatics", "biomolecular MD" ], "properties": [ "energy", "forces" ], "frameworks": [ "PyTorch" ], "maintenance": "experimental", "lastReviewed": "2026-05-05", "trainingData": [ "custom periodic + biomolecular DFT" ], "tags": [ "SO(3)-equivariant", "long-range-electrostatics", "reciprocal space", "Ewald" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "—", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": true }, { "id": "allscaip", "type": "node", "category": "Transformer", "label": "AllScAIP", "year": 2026, "author": "Qu, Wood, Krishnapriyan, Ulissi (FAIR / Meta, UC Berkeley, LBNL)", "x": 2620, "y": 320, "desc": "Scalable, energy-conserving, attention-based MLIP that pairs local neighborhood self-attention with a global all-to-all node attention layer in which every atom attends to every other atom. The data-driven all-to-all component captures long-range interactions without explicit electrostatic priors and remains the most durable ingredient as data and model size scale to O(100M) training samples. Sits atop the OMol25 leaderboard at release while remaining competitive on OMat24 (materials) and OC20 (catalysts); cuts long-range distance-scaling error by ~90% versus the next-best foundation model, with stable long-timescale MD recovering experimental densities and heats of vaporisation.", "githubUrl": "https://github.com/facebookresearch/fairchem", "paperUrl": "https://arxiv.org/abs/2603.06567", "isNew": true, "coverage": [ "organic molecules", "general materials", "catalysts" ], "useCases": [ "large-scale MD", "long-range chemistry" ], "properties": [ "energy", "forces" ], "frameworks": [ "ASE", "PyTorch" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-05", "trainingData": [ "OMol25", "OMat24", "OC20" ], "tags": [ "transformer", "all-to-all attention", "long-range-electrostatics", "foundation model", "charge-aware", "spin-aware" ], "supportsCharges": true, "supportsSpins": true, "elementsCovered": "all elements covered by OMol25 / OMat24 (~89 elements)", "equivariance": "learnt", "architecture": "gnn", "usesAttention": true, "longRange": true, "trainingSetSize": 100000000 }, { "id": "mace_mag", "type": "node", "category": "Equivariant", "label": "MACE-Magnetic", "year": 2026, "author": "Ho, van der Oord, Darby, Csányi, Ortner et al. (Cambridge / UBC / ACEsuit)", "x": 2900, "y": 150, "desc": "Equivariant many-body message-passing interatomic potential extending the MACE framework to magnetic materials by embedding atomic magnetic moments as explicit degrees of freedom alongside positions. Learns physically consistent and transferable representations of magnetic behaviour beyond collinear approximations and can incorporate spin-orbit coupling, achieving near density-functional-theory accuracy with strong data efficiency by fine-tuning from a pre-trained foundation model. Targets structural transformations, finite-temperature magnetic phenomena, and high-throughput screening of strongly spin-orbit coupled materials.", "paperUrl": "https://arxiv.org/abs/2604.08143", "isNew": true, "coverage": [ "magnetic materials", "strongly spin-orbit coupled materials" ], "useCases": [ "magnetic structure prediction", "finite-temperature magnetism" ], "properties": [ "energy", "forces", "magnetic_moment" ], "frameworks": [ "ASE", "LAMMPS" ], "maintenance": "experimental", "lastReviewed": "2026-05-05", "trainingData": [ "custom magnetic materials DFT" ], "tags": [ "equivariant", "MACE", "magnetic moments", "spin-orbit coupling" ], "supportsCharges": false, "supportsSpins": true, "elementsCovered": "—", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "allegro_moe", "type": "node", "category": "Equivariant", "label": "Allegro-MoE", "year": 2026, "author": "Nascimento, Descoteaux, Zichi, Tan, Witt, Molinari, Mantha, Kitchaev, Kornbluth, Gadelrab, Tuffile, Kozinsky (Harvard / Bosch)", "x": 2900, "y": 320, "desc": "Multifidelity Mixture-of-Experts framework built on the strictly local E(3)-equivariant Allegro architecture. Spatially partitions the simulation domain into chemically complex regions (e.g. reactive interfaces) and simple regions (e.g. bulk lattices) and assigns Allegro experts of different capacity to each, enabling expensive high-fidelity inference only where required while a cheaper expert handles the rest of the cell.", "paperUrl": "https://arxiv.org/abs/2604.26143", "isNew": true, "coverage": [ "heterogeneous interfaces", "reactive interfaces", "bulk lattices" ], "useCases": [ "multifidelity MD", "large heterogeneous-system simulation" ], "properties": [ "energy", "forces" ], "frameworks": [ "ASE", "LAMMPS", "PyTorch" ], "maintenance": "experimental", "lastReviewed": "2026-05-05", "trainingData": [ "custom heterogeneous-system DFT" ], "tags": [ "equivariant", "Allegro", "mixture-of-experts", "multifidelity", "spatial partitioning" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "—", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "hi_mlip", "type": "node", "category": "Equivariant", "label": "Hi-MLIP", "year": 2026, "author": "Yin, Ouyang, Fan, Lin, Hu, Lv, Cao, Xiao, Chen, Ren (ByteDance Seed / Tsinghua / Peking University)", "x": 3180, "y": 150, "desc": "Hessian-informed machine learning interatomic potential trained with the Hessian-INformed Training (HINT) protocol — Hessian pre-training, configuration sampling, curriculum learning, and a stochastic projected Hessian loss — to attain Hessian-level accuracy with two to four orders of magnitude fewer high-fidelity Hessian labels than standard training. Substantially improves transition-state search and brings Gibbs free-energy predictions close to chemical accuracy in data-scarce regimes, and reproduces phonon renormalization and superconducting Tc of strongly anharmonic hydrides in close agreement with experiment.", "paperUrl": "https://arxiv.org/abs/2603.25373", "isNew": true, "coverage": [ "transition states", "anharmonic hydrides", "phonon systems" ], "useCases": [ "transition-state search", "Gibbs free energy", "phonon prediction" ], "properties": [ "energy", "forces" ], "frameworks": [ "PyTorch" ], "maintenance": "experimental", "lastReviewed": "2026-05-05", "trainingData": [ "custom Hessian-labeled DFT" ], "tags": [ "Hessian", "HINT", "transition states", "anharmonic", "curvature-aware" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "—", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "spookynet", "type": "node", "category": "Transformer", "label": "SpookyNet", "year": 2021, "author": "Unke, Chmiela, Gastegger, Schütt, Sauceda, Müller", "x": 3460, "y": 150, "desc": "MLIP that explicitly captures electronic degrees of freedom and nonlocal effects, modelled via self-attention in a transformer architecture. Total charge and spin multiplicity are injected as global tokens that condition the local atomic embeddings — a direct conceptual ancestor of charge/spin-conditioned foundation models such as UMA and MACE-POLAR-1.", "githubUrl": "https://github.com/OUnke/SpookyNet", "paperUrl": "https://www.nature.com/articles/s41467-021-27504-0", "coverage": [ "organic molecules", "charged molecules" ], "useCases": [ "charge/spin-conditioned MD", "ionic systems" ], "properties": [ "energy", "forces", "dipole" ], "frameworks": [ "PyTorch", "ASE" ], "license": "MIT", "maintenance": "archived", "lastReviewed": "2026-05-05", "trainingData": [ "custom DFT", "QMspin", "QM7-X" ], "tags": [ "transformer", "self-attention", "charge-aware", "spin-aware", "long-range-electrostatics" ], "supportsCharges": true, "supportsSpins": true, "elementsCovered": "—", "equivariance": "invariant", "architecture": "gnn", "usesAttention": true, "longRange": true }, { "id": "gems", "type": "node", "category": "Transformer", "label": "GEMS", "year": 2024, "author": "Unke, Stöhr, Ganscha, Unterthiner, Maennel, Kashubin, Ahlin, Gastegger, Sandonas, Berryman, Tkatchenko, Müller", "x": 3460, "y": 320, "desc": "SpookyNet-based biomolecular force-field framework that combines top-down (whole-protein) and bottom-up (fragment) sampling to train transferable, quantum-accurate ML potentials for proteins and condensed-phase biomolecular dynamics.", "paperUrl": "https://www.science.org/doi/10.1126/sciadv.adn4397", "coverage": [ "proteins", "biomolecules" ], "useCases": [ "biomolecular MD", "protein simulation" ], "properties": [ "energy", "forces" ], "frameworks": [ "PyTorch" ], "license": "proprietary", "maintenance": "experimental", "lastReviewed": "2026-05-05", "trainingData": [ "custom protein/biomolecular DFT" ], "tags": [ "transformer", "biomolecular", "fragment sampling", "charge-aware" ], "supportsCharges": true, "supportsSpins": true, "elementsCovered": "—", "equivariance": "invariant", "architecture": "gnn", "usesAttention": true, "longRange": true }, { "id": "bpnn", "type": "node", "category": "Descriptor", "label": "BPNN", "year": 2007, "author": "Behler & Parrinello", "x": 100, "y": 650, "desc": "The Behler–Parrinello high-dimensional neural network potential — the 2007 paper that introduced symmetry functions and the atomic-decomposition framework underlying essentially every modern MLIP.", "githubUrl": "https://github.com/CompPhysVienna/n2p2", "paperUrl": "https://doi.org/10.1103/PhysRevLett.98.146401", "coverage": [ "general materials" ], "useCases": [ "high-dimensional NN potential", "MD simulation" ], "properties": [ "energy", "forces" ], "frameworks": [ "LAMMPS", "ASE" ], "license": "GPL-3.0-or-later", "maintenance": "maintained", "lastReviewed": "2026-05-05", "lastUpdated": "2024-11-22", "trainingData": [ "custom DFT" ], "tags": [ "descriptor", "atomic NN", "symmetry functions", "BPNN", "HDNNP" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "—", "equivariance": "invariant", "architecture": "descriptor", "usesAttention": false, "longRange": false }, { "id": "ace", "type": "node", "category": "Descriptor", "label": "ACE", "year": 2019, "author": "Drautz (ICAMS)", "x": 380, "y": 650, "desc": "Atomic Cluster Expansion: a complete, systematically improvable many-body basis for the local atomic environment; the mathematical backbone of PACE/GRACE and a strong influence on MACE.", "githubUrl": "https://github.com/ICAMS/lammps-user-pace", "paperUrl": "https://arxiv.org/abs/1902.10301", "coverage": [ "general materials" ], "useCases": [ "many-body MLIP basis", "production MD" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "LAMMPS", "ASE" ], "license": "Apache-2.0", "maintenance": "maintained", "lastReviewed": "2026-05-05", "lastUpdated": "2026-02-09", "trainingData": [ "custom DFT" ], "tags": [ "descriptor", "ACE", "many-body", "linear basis" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "—", "equivariance": "invariant", "architecture": "descriptor", "usesAttention": false, "longRange": false }, { "id": "gap", "type": "node", "category": "Descriptor", "label": "GAP", "year": 2010, "author": "Bartók, Payne, Kondor, Csányi (Cambridge)", "x": 100, "y": 550, "desc": "Gaussian Approximation Potentials: a kernel-based MLIP fit to ab-initio energies and forces using SOAP-like many-body descriptors; the canonical kernel reference for descriptor-based potentials.", "githubUrl": "https://github.com/libAtoms/QUIP", "paperUrl": "https://doi.org/10.1103/PhysRevLett.104.136403", "coverage": [ "general materials", "amorphous systems" ], "useCases": [ "kernel MLIP fitting", "amorphous solids MD" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "LAMMPS", "ASE" ], "license": "GPL-3.0-or-later", "maintenance": "maintained", "lastReviewed": "2026-05-05", "lastUpdated": "2026-04-01", "trainingData": [ "custom DFT" ], "tags": [ "descriptor", "kernel", "SOAP", "Gaussian process" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "—", "equivariance": "invariant", "architecture": "descriptor", "usesAttention": false, "longRange": false }, { "id": "deepmd", "type": "node", "category": "Descriptor", "label": "DeepMD", "year": 2018, "author": "DeepModeling", "x": 380, "y": 550, "desc": "Deep Potential Molecular Dynamics: local frame descriptors + deep networks giving ab-initio-level accuracy with linear scaling, widely used for large MD.", "githubUrl": "https://github.com/deepmodeling/deepmd-kit", "paperUrl": "https://doi.org/10.1016/j.cpc.2018.02.017", "coverage": [ "general materials", "molecules" ], "useCases": [ "large-scale ab-initio MD", "GPU MD" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "LAMMPS", "ASE", "PyTorch" ], "license": "LGPL-3.0", "maintenance": "active", "lastReviewed": "2026-05-05", "lastUpdated": "2026-03-19", "trainingData": [ "custom DFT" ], "tags": [ "descriptor", "deep potential", "local frame" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "—", "equivariance": "invariant", "architecture": "descriptor", "usesAttention": false, "longRange": false }, { "id": "dpa2", "type": "node", "category": "Descriptor", "label": "DPA-2", "year": 2024, "author": "DeepModeling", "x": 660, "y": 550, "desc": "Second-generation Deep Potential architecture with attention and multi-task heads, targeting a universal deep potential for diverse chemistries.", "githubUrl": "https://github.com/deepmodeling/deepmd-kit", "paperUrl": "https://arxiv.org/abs/2312.15492", "coverage": [ "general materials", "molecules", "diverse chemistries" ], "useCases": [ "universal deep potential", "multi-task fine-tuning" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "LAMMPS", "ASE", "PyTorch" ], "license": "LGPL-3.0", "maintenance": "active", "lastReviewed": "2026-05-05", "lastUpdated": "2026-03-19", "trainingData": [ "multi-task DFT" ], "tags": [ "descriptor", "attention", "multi-task", "foundation model" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "—", "equivariance": "invariant", "architecture": "descriptor", "usesAttention": true, "longRange": false }, { "id": "dpa3", "type": "node", "category": "Descriptor", "label": "DPA-3", "year": 2025, "author": "DeepModeling", "x": 1510, "y": 550, "desc": "Message-passing graph neural network built on a Line Graph Series (LiGS) that updates bond, angle, and dihedral representations while preserving energy conservation and physical symmetries; designed for Large Atomistic Models with clean scaling in model size, data, and compute. The DPA-3.1-3M variant trained on OpenLAM-v1 tops zero-shot generalization across 12 downstream tasks.", "githubUrl": "https://github.com/deepmodeling/deepmd-kit", "paperUrl": "https://arxiv.org/abs/2506.01686", "isNew": true, "coverage": [ "general materials", "molecules" ], "useCases": [ "large atomistic models", "zero-shot generalization" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "LAMMPS", "ASE", "PyTorch" ], "license": "LGPL-3.0", "maintenance": "active", "lastReviewed": "2026-05-05", "lastUpdated": "2026-03-19", "trainingData": [ "OpenLAM-v1" ], "tags": [ "descriptor", "message-passing", "line graph", "foundation model" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "—", "architecture": "descriptor", "equivariance": "invariant", "usesAttention": true, "longRange": false }, { "id": "mattersim", "type": "node", "category": "Equivariant", "label": "MatterSim", "year": 2024, "author": "Microsoft", "x": 950, "y": 550, "desc": "Microsoft foundation MLIP using a Graphormer transformer backbone with explicit translation/periodic-boundary invariance and equivariant features for materials. Trained on large-scale ab-initio data spanning 0-5000 K and pressures up to 1000 GPa as a reusable simulator for materials discovery and high-throughput computation.", "githubUrl": "https://github.com/microsoft/mattersim", "paperUrl": "https://arxiv.org/abs/2405.04967", "coverage": [ "general materials" ], "useCases": [ "materials discovery", "high-throughput screening" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE", "PyTorch" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-05", "trainingData": [ "custom MatterSim DFT" ], "tags": [ "transformer", "foundation model" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements up to Z=89", "equivariance": "constrained", "architecture": "gnn", "usesAttention": true, "longRange": false, "numElements": 89 }, { "id": "nep89", "type": "node", "category": "Descriptor", "label": "NEP89", "year": 2025, "author": "Chalmers (Fan group) / Liang et al.", "x": 1230, "y": 550, "desc": "Universal neuroevolution-potential foundation model spanning 89 elements across inorganic and organic materials, trained via separable natural evolution strategies and distributed in GPUMD for empirical-potential-like speed.", "githubUrl": "https://github.com/brucefan1983/GPUMD", "paperUrl": "https://arxiv.org/abs/2504.21286", "isNew": true, "coverage": [ "general materials", "organic molecules" ], "useCases": [ "high-throughput MD", "empirical-potential-speed simulation" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "LAMMPS" ], "license": "GPL-3.0", "maintenance": "active", "lastReviewed": "2026-05-05", "trainingData": [ "custom 89-element DFT" ], "tags": [ "descriptor", "neuroevolution", "foundation model" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "89 elements", "equivariance": "invariant", "architecture": "descriptor", "usesAttention": false, "longRange": false, "numElements": 89 }, { "id": "liten", "type": "node", "category": "Equivariant", "label": "LiTEN-FF", "year": 2026, "author": "Zhejiang Univ. (Hou lab) / Su et al.", "x": 1230, "y": 650, "desc": "Equivariant network with Tensorized Quadrangle Attention (TQA) that captures three- and four-body interactions in linear time; pre-trained on nablaDFT and fine-tuned on SPICE as a quantum-accurate biomolecular force-field foundation model with ~10x faster inference than MACE-OFF.", "githubUrl": "https://github.com/lingcon01/LiTEN", "paperUrl": "https://arxiv.org/abs/2507.00884", "isNew": true, "coverage": [ "biomolecules", "organic molecules" ], "useCases": [ "biomolecular force field", "drug discovery" ], "properties": [ "energy", "forces" ], "frameworks": [ "ASE", "PyTorch" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-05", "trainingData": [ "nablaDFT", "SPICE" ], "tags": [ "equivariant", "linear tensor", "biomolecular", "attention" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "SPICE / nablaDFT organic and biomolecular chemistry", "equivariance": "constrained", "architecture": "gnn", "usesAttention": true, "longRange": false }, { "id": "matris", "type": "node", "category": "Invariant", "label": "MatRIS", "year": 2026, "author": "Zhou et al. (CAS / UCAS)", "x": 1510, "y": 650, "desc": "Invariant foundation MLIP using a separable O(N) attention mechanism for three-body interactions; 10M-parameter models trained on OMat24 / MPTrj / sAlex match equivariant SOTA on Matbench-Discovery (F1 0.847) at >13x lower training cost than eSEN-30M-MP.", "githubUrl": "https://github.com/HPC-AI-Team/MatRIS", "paperUrl": "https://arxiv.org/abs/2603.02002", "isNew": true, "coverage": [ "general materials" ], "useCases": [ "foundation MLIP", "Matbench-Discovery" ], "properties": [ "energy", "forces", "stress", "magnetic_moment" ], "frameworks": [ "ASE", "PyTorch" ], "license": "BSD-3-Clause", "maintenance": "active", "lastReviewed": "2026-05-05", "trainingData": [ "OMat24", "MPTrj", "sAlex" ], "tags": [ "invariant", "attention", "foundation model" ], "supportsCharges": false, "supportsSpins": true, "elementsCovered": "all elements covered by OMat24 / MPTrj (~89 elements)", "equivariance": "invariant", "architecture": "gnn", "usesAttention": true, "longRange": false, "numElements": 89 }, { "id": "matris_moe", "type": "node", "category": "Invariant", "label": "MatRIS-MoE", "year": 2026, "author": "Zhou et al. (CAS / ICT, HPC-AI-Team)", "x": 1790, "y": 650, "desc": "Billion-parameter Mixture-of-Experts extension of MatRIS that inserts sparse expert modules around the self-attention layer — a message-update MoE for message construction and a feature-update MoE for post-attention refinement — with element-type routing that keeps the activated expert set time-independent and the potential energy surface continuous. Released in M (2.47B) and L (11.50B) variants and trained on heterogeneous domains (molecules, materials, catalysis, MOFs, and direct air capture) via the new Janus hybrid-parallel framework, attaining 1.0–1.2 EFLOPS at >90% parallel efficiency and compressing billion-parameter uMLIP training from weeks to hours on Exascale supercomputers.", "githubUrl": "https://github.com/HPC-AI-Team/MatRIS", "paperUrl": "https://arxiv.org/abs/2604.15821", "isNew": true, "coverage": [ "general materials", "molecules", "catalysts", "MOFs", "direct air capture" ], "useCases": [ "billion-parameter MLIP", "exascale training", "multi-task foundation model" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "PyTorch" ], "license": "BSD-3-Clause", "maintenance": "active", "lastReviewed": "2026-05-05", "trainingData": [ "OMat24", "MPTrj", "sAlex", "OMol25", "OC20", "ODAC23" ], "tags": [ "invariant", "mixture-of-experts", "billion-parameter", "multi-task", "foundation model" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements covered by OMat24 / MPTrj / OMol25 (~89 elements)", "equivariance": "invariant", "architecture": "gnn", "usesAttention": true, "longRange": false, "numElements": 89 }, { "id": "snap", "type": "node", "category": "Descriptor", "label": "SNAP", "year": 2015, "author": "Thompson, Swiler, Trott, Foiles, Tucker (Sandia)", "x": 1790, "y": 550, "desc": "Spectral Neighbor Analysis Potential: a linear descriptor MLIP fit to bispectrum components of the local atomic environment, designed for high-throughput LAMMPS simulations. The canonical industrial-scale linear descriptor potential alongside GAP and MTP.", "paperUrl": "https://doi.org/10.1016/j.jcp.2014.12.018", "coverage": [ "general materials", "metals" ], "useCases": [ "industrial-scale linear MLIP", "high-throughput MD" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "LAMMPS" ], "license": "GPL-2.0", "maintenance": "maintained", "lastReviewed": "2026-05-05", "trainingData": [ "custom DFT" ], "tags": [ "descriptor", "linear", "bispectrum" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "—", "equivariance": "invariant", "architecture": "descriptor", "usesAttention": false, "longRange": false }, { "id": "mtp", "type": "node", "category": "Descriptor", "label": "MTP", "year": 2016, "author": "Shapeev (Skoltech)", "x": 2060, "y": 550, "desc": "Moment Tensor Potentials: a systematically improvable linear MLIP based on contractions of moment tensors of the local environment. Pairs naturally with D-optimality / MaxVol active learning (Podryabinkin & Shapeev 2017), making the MTP family the canonical reference for AL-native potential fitting.", "githubUrl": "https://gitlab.com/ashapeev/mlip-2", "paperUrl": "https://arxiv.org/abs/1512.06054", "coverage": [ "general materials", "metals", "alloys" ], "useCases": [ "active learning MLIP fitting", "high-throughput MD" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "LAMMPS" ], "license": "BSD-2-Clause", "maintenance": "maintained", "lastReviewed": "2026-05-05", "trainingData": [ "custom DFT" ], "tags": [ "descriptor", "linear", "moment tensors", "active-learning" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "—", "equivariance": "invariant", "architecture": "descriptor", "usesAttention": false, "longRange": false }, { "id": "sgdml", "type": "node", "category": "Descriptor", "label": "sGDML", "year": 2017, "author": "Chmiela, Sauceda, Poltavsky, Müller, Tkatchenko", "x": 2340, "y": 550, "desc": "Symmetric Gradient-Domain Machine Learning: a kernel ridge regression force field fit directly in the gradient domain, with energy obtained by closed-form integration. Strong on small molecules where rigorous symmetry handling is critical.", "githubUrl": "https://github.com/stefanch/sGDML", "paperUrl": "https://www.science.org/doi/10.1126/sciadv.1603015", "coverage": [ "small organic molecules" ], "useCases": [ "high-accuracy molecular MD", "MD17/MD22 benchmarks" ], "properties": [ "energy", "forces" ], "frameworks": [ "ASE", "PyTorch" ], "license": "MIT", "maintenance": "maintained", "lastReviewed": "2026-05-05", "trainingData": [ "MD17", "MD22" ], "tags": [ "descriptor", "kernel", "gradient-domain" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "—", "equivariance": "invariant", "architecture": "descriptor", "usesAttention": false, "longRange": false }, { "id": "flare", "type": "node", "category": "Descriptor", "label": "FLARE", "year": 2020, "author": "Vandermause, Torrisi, Batzner, Sun, Kolpak, Kozinsky (Harvard)", "x": 2620, "y": 550, "desc": "Fast Learning of Atomistic Rare Events: a Gaussian-process-regression Bayesian potential trained on-the-fly during MD, with GP-uncertainty driving when to call DFT vs. trust the surrogate. Includes tabulated/mapped force-field export for production-speed MD; principal kernel-based AL framework and immediate predecessor of the Kozinsky group's NequIP/Allegro line.", "githubUrl": "https://github.com/mir-group/flare", "paperUrl": "https://www.nature.com/articles/s41524-020-0283-z", "coverage": [ "general materials", "rare events" ], "useCases": [ "on-the-fly active learning", "Bayesian uncertainty MD" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE", "LAMMPS" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-05", "trainingData": [ "custom DFT" ], "tags": [ "descriptor", "Gaussian process", "active-learning", "Bayesian" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "—", "equivariance": "invariant", "architecture": "descriptor", "usesAttention": false, "longRange": false }, { "id": "aimnet", "type": "node", "category": "Descriptor", "label": "AIMNet", "year": 2019, "author": "Zubatyuk, Smith, Leszczynski, Isayev", "x": 2060, "y": 650, "desc": "Original Atoms-in-Molecules Network: a self-consistent message-passing potential that propagates atomic environment vectors through repeated neighbour updates, learning charge-aware atomic representations for organic chemistry. The conceptual precursor to AIMNet-NSE and AIMNet2.", "githubUrl": "https://github.com/aiqm/aimnet", "paperUrl": "https://pubs.rsc.org/en/content/articlehtml/2019/sc/c9sc00531e", "coverage": [ "organic molecules" ], "useCases": [ "organic chemistry MD", "atomic charge prediction" ], "properties": [ "energy", "forces" ], "frameworks": [ "ASE", "PyTorch" ], "license": "MIT", "maintenance": "archived", "lastReviewed": "2026-05-05", "trainingData": [ "ANI-1x" ], "tags": [ "descriptor", "message-passing", "AIM" ], "supportsCharges": true, "supportsSpins": false, "elementsCovered": "—", "equivariance": "invariant", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "aimnet_nse", "type": "node", "category": "Descriptor", "label": "AIMNet-NSE", "year": 2021, "author": "Zubatyuk, Smith, Nebgen, Tretiak, Isayev", "x": 2340, "y": 650, "desc": "Neural Spin Equilibration variant of AIMNet that handles arbitrary total charge and spin multiplicity through an SCF-like message-passing loop, predicting transferable atomic charges and spins. The conceptual ancestor of charge/spin-conditioned models like AIMNet2 and SpookyNet.", "githubUrl": "https://github.com/zubatyuk/aimnet-nse", "paperUrl": "https://www.nature.com/articles/s41467-021-24904-0", "coverage": [ "organic molecules", "charged species" ], "useCases": [ "charge/spin equilibration", "ionic + radical chemistry" ], "properties": [ "energy", "forces" ], "frameworks": [ "PyTorch" ], "license": "MIT", "maintenance": "archived", "lastReviewed": "2026-05-05", "trainingData": [ "custom organic DFT" ], "tags": [ "descriptor", "message-passing", "charge-aware", "spin-aware", "SCF" ], "supportsCharges": true, "supportsSpins": true, "elementsCovered": "—", "equivariance": "invariant", "architecture": "gnn", "usesAttention": false, "longRange": true }, { "id": "hip_nn", "type": "node", "category": "Invariant", "label": "HIP-NN", "year": 2018, "author": "Lubbers, Smith, Barros (LANL)", "x": 2620, "y": 650, "desc": "Hierarchically Interacting Particle Neural Network: a message-passing potential that decomposes atomic energies into a hierarchy of n-body terms with explicit residual structure, providing both interpretable hierarchical decomposition and natural uncertainty estimates from the residual stack.", "githubUrl": "https://github.com/lanl/hippynn", "paperUrl": "https://pubs.aip.org/aip/jcp/article/148/24/241715/962591", "coverage": [ "general materials", "molecules" ], "useCases": [ "hierarchical decomposition", "uncertainty quantification" ], "properties": [ "energy", "forces" ], "frameworks": [ "ASE", "LAMMPS", "PyTorch" ], "license": "BSD-3-Clause", "maintenance": "active", "lastReviewed": "2026-05-05", "trainingData": [ "custom DFT" ], "tags": [ "invariant", "message-passing", "hierarchical", "uncertainty" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "—", "equivariance": "invariant", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "ani", "type": "node", "category": "Descriptor", "label": "ANI-2x", "year": 2020, "author": "Devereux, Smith, Huddleston, Barros, Zubatyuk, Isayev, Roitberg", "x": 100, "y": 750, "desc": "ANI variant extending coverage to seven elements (H, C, N, O, S, F, Cl); widely used in drug discovery for fast geometry and energy scans on organic chemistry.", "githubUrl": "https://github.com/aiqm/torchani", "paperUrl": "https://pubs.acs.org/doi/10.1021/acs.jctc.0c00121", "coverage": [ "organic molecules", "drug-like molecules" ], "useCases": [ "drug discovery", "geometry/energy scans" ], "properties": [ "energy", "forces" ], "frameworks": [ "ASE", "PyTorch" ], "license": "MIT", "maintenance": "maintained", "lastReviewed": "2026-05-05", "trainingData": [ "ANI-2x" ], "tags": [ "descriptor", "atomic NN", "symmetry functions" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "H, C, N, O, S, F, Cl", "equivariance": "invariant", "architecture": "descriptor", "usesAttention": false, "longRange": false, "numElements": 7 }, { "id": "ani_1", "type": "node", "category": "Descriptor", "label": "ANI-1", "year": 2017, "author": "Smith, Isayev, Roitberg", "x": 2340, "y": 750, "desc": "Original ANI: an atomistic neural network potential built on Behler-style symmetry function descriptors and per-element atomic networks, trained on ~22M DFT structures of organic molecules covering H, C, N, O.", "githubUrl": "https://github.com/aiqm/torchani", "paperUrl": "https://pubs.rsc.org/en/content/articlehtml/2017/sc/c6sc05720a", "coverage": [ "organic molecules" ], "useCases": [ "fast organic chemistry energies" ], "properties": [ "energy", "forces" ], "frameworks": [ "ASE", "PyTorch" ], "license": "MIT", "maintenance": "maintained", "lastReviewed": "2026-05-05", "trainingData": [ "ANI-1" ], "tags": [ "descriptor", "atomic NN", "symmetry functions" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "H, C, N, O", "equivariance": "invariant", "architecture": "descriptor", "usesAttention": false, "longRange": false, "numElements": 4 }, { "id": "ani_1x", "type": "node", "category": "Descriptor", "label": "ANI-1x", "year": 2018, "author": "Smith, Nebgen, Lubbers, Isayev, Roitberg", "x": 2620, "y": 750, "desc": "Active-learned extension of ANI-1: applies query-by-committee active learning to grow the training set towards a chemically diverse organic-molecule sampling, dramatically improving generalisation while maintaining ANI-style efficiency.", "githubUrl": "https://github.com/aiqm/torchani", "paperUrl": "https://pubs.aip.org/aip/jcp/article/148/24/241733/962534", "coverage": [ "organic molecules" ], "useCases": [ "active learning", "diverse chemistry" ], "properties": [ "energy", "forces" ], "frameworks": [ "ASE", "PyTorch" ], "license": "MIT", "maintenance": "maintained", "lastReviewed": "2026-05-05", "trainingData": [ "ANI-1x" ], "tags": [ "descriptor", "atomic NN", "active-learning" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "H, C, N, O", "equivariance": "invariant", "architecture": "descriptor", "usesAttention": false, "longRange": false, "numElements": 4 }, { "id": "ani_1ccx", "type": "node", "category": "Descriptor", "label": "ANI-1ccx", "year": 2020, "author": "Smith, Nebgen, Zubatyuk, Lubbers, Devereux, Barros, Tretiak, Isayev, Roitberg", "x": 2900, "y": 750, "desc": "ANI-1x transfer-learned to a CCSD(T)/CBS-quality reference, yielding a near-coupled-cluster-accuracy organic-chemistry potential at deep-network speed. The first widely deployed example of using transfer learning to push beyond DFT.", "githubUrl": "https://github.com/aiqm/torchani", "paperUrl": "https://www.nature.com/articles/s41467-019-10827-4", "coverage": [ "organic molecules" ], "useCases": [ "coupled-cluster-level chemistry", "drug discovery" ], "properties": [ "energy", "forces" ], "frameworks": [ "ASE", "PyTorch" ], "license": "MIT", "maintenance": "maintained", "lastReviewed": "2026-05-05", "trainingData": [ "ANI-1ccx", "ANI-1x" ], "tags": [ "descriptor", "atomic NN", "transfer learning", "CCSD(T)" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "H, C, N, O", "equivariance": "invariant", "architecture": "descriptor", "usesAttention": false, "longRange": false, "numElements": 4 }, { "id": "aimnet2", "type": "node", "category": "Descriptor", "label": "AIMNet2", "year": 2025, "author": "Anstine, Zubatyuk & Isayev (CMU)", "x": 1510, "y": 750, "desc": "Second-generation Atoms-in-Molecules Network potential covering 14 elements (H, B, C, N, O, F, Si, P, S, Cl, As, Se, Br, I) in neutral and charged states; combines ML-parameterised short-range terms with physics-based long-range electrostatics, trained on ~20M hybrid-DFT (wB97M-D3) calculations for routine use as a DFT replacement in organic and elemental-organic chemistry.", "githubUrl": "https://github.com/isayevlab/aimnetcentral", "paperUrl": "https://doi.org/10.1039/D4SC08572H", "isNew": true, "coverage": [ "organic molecules", "elemental-organic chemistry", "charged molecules" ], "useCases": [ "drug discovery", "DFT replacement" ], "properties": [ "energy", "forces" ], "frameworks": [ "ASE", "PyTorch" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-05", "trainingData": [ "custom hybrid-DFT (wB97M-D3)" ], "tags": [ "descriptor", "AIM", "charge-aware", "long-range-electrostatics" ], "supportsCharges": true, "supportsSpins": false, "elementsCovered": "H, B, C, N, O, F, Si, P, S, Cl, As, Se, Br, I", "equivariance": "invariant", "architecture": "gnn", "usesAttention": false, "longRange": true, "trainingSetSize": 20000000, "numElements": 14 }, { "id": "aceff", "type": "node", "category": "Equivariant", "label": "AceFF", "year": 2026, "author": "Farr, Doerr, Mirarchi, Sabanés Zariquiey & De Fabritiis (Acellera Labs / UPF Barcelona)", "x": 2060, "y": 750, "desc": "Drug-discovery-oriented MLIP built on the TensorNet2 architecture — a refined vector–scalar equivariant TensorNet that adds scalar partial-charge features, performs neutral charge equilibration, and includes a long-range Coulomb energy term. Pretrained on a large dataset of drug-like compounds covering H, B, C, N, O, F, Si, P, S, Cl, Br, I in neutral and charged states; balances DFT-level accuracy on torsion scans, MD trajectories, and batched minimisations with high-throughput inference suitable for FEP and lead-optimisation workflows.", "githubUrl": "https://huggingface.co/Acellera/AceFF-2.0", "paperUrl": "https://arxiv.org/abs/2601.00581", "isNew": true, "coverage": [ "drug-like molecules", "organic molecules" ], "useCases": [ "FEP", "lead optimisation", "drug discovery" ], "license": "Apache-2.0", "maintenance": "active", "lastReviewed": "2026-05-05", "properties": [ "energy", "forces" ], "frameworks": [ "ASE", "PyTorch" ], "trainingData": [ "custom drug-like compound DFT" ], "tags": [ "TensorNet2", "drug discovery", "charge-aware", "long-range-electrostatics" ], "supportsCharges": true, "supportsSpins": false, "elementsCovered": "H, B, C, N, O, F, Si, P, S, Cl, Br, I", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": true, "numElements": 12 }, { "id": "schnet", "type": "node", "category": "Invariant", "label": "SchNet", "year": 2017, "author": "Schütt et al.", "x": 380, "y": 750, "desc": "Continuous-filter convolutional network that introduced smooth, translation-invariant filters for molecules; the baseline for many later invariant message-passing potentials. Predicts energies and forces with all-atom symmetry preserved.", "githubUrl": "https://github.com/atomistic-machine-learning/schnetpack", "paperUrl": "https://arxiv.org/abs/1706.08566", "coverage": [ "organic molecules", "general materials" ], "useCases": [ "GNN MLIP baseline", "molecular property prediction" ], "properties": [ "energy", "forces" ], "frameworks": [ "ASE", "PyTorch" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-05", "lastUpdated": "2025-12-19", "trainingData": [ "QM9", "MD17" ], "tags": [ "invariant", "message-passing", "continuous-filter convolution" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "dataset-dependent", "equivariance": "invariant", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "dimenet", "type": "node", "category": "Invariant", "label": "DimeNet++", "year": 2020, "author": "Gasteiger et al.", "x": 660, "y": 750, "desc": "Directional message passing network with spherical basis functions that explicitly encode bond angles, improving data efficiency over SchNet-style models. The original paper (arXiv:2003.03123) introduced the directional MP framework; the follow-up DimeNet++ (arXiv:2011.14115) refined it with faster, uncertainty-aware variants.", "githubUrl": "https://github.com/gasteigerjo/dimenet", "paperUrl": "https://arxiv.org/abs/2003.03123", "coverage": [ "organic molecules" ], "useCases": [ "molecular property prediction", "data-efficient GNN" ], "properties": [ "energy", "forces" ], "frameworks": [ "PyTorch" ], "license": "Hippocratic-2.1", "maintenance": "archived", "lastReviewed": "2026-05-05", "trainingData": [ "QM9", "MD17" ], "tags": [ "invariant", "message-passing", "directional", "spherical" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "dataset-dependent", "equivariance": "invariant", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "gemnet", "type": "node", "category": "Equivariant", "label": "GemNet-OC", "year": 2021, "author": "TUM / OC20", "x": 950, "y": 750, "desc": "High-capacity spherical message-passing architecture used in the OC20/OC22 benchmarks; very strong for catalyst adsorption and surface chemistry.", "githubUrl": "https://github.com/OpenCatalystProject/ocp", "paperUrl": "https://arxiv.org/abs/2106.08903", "coverage": [ "catalysts", "surfaces", "oxides" ], "useCases": [ "catalyst adsorption", "surface chemistry" ], "properties": [ "energy", "forces" ], "frameworks": [ "PyTorch", "ASE" ], "license": "MIT", "maintenance": "maintained", "lastReviewed": "2026-05-05", "trainingData": [ "OC20", "OC22" ], "tags": [ "equivariant", "message-passing", "spherical" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements covered by OC20 / OC22 (~56 elements)", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false, "numElements": 56 }, { "id": "painn", "type": "node", "category": "Equivariant", "label": "PaiNN", "year": 2021, "author": "Schütt et al.", "x": 380, "y": 900, "desc": "Polarizable Atom Interaction Neural Network: uses coupled scalar/vector features to capture forces and dipoles with E(3) equivariance at lower cost than full tensors.", "githubUrl": "https://github.com/atomistic-machine-learning/schnetpack", "paperUrl": "https://arxiv.org/abs/2102.03150", "coverage": [ "organic molecules" ], "useCases": [ "dipole prediction", "polarizable response" ], "properties": [ "energy", "forces", "dipole", "polarizability" ], "frameworks": [ "ASE", "PyTorch" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-05", "lastUpdated": "2025-12-19", "trainingData": [ "QM9", "MD17" ], "tags": [ "equivariant", "vector", "scalar", "polarizable" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "dataset-dependent", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "sevennet", "type": "node", "category": "Equivariant", "label": "SevenNet", "year": 2024, "author": "Seoul Nat. Univ.", "x": 660, "y": 900, "desc": "NequIP-based scalable equivariant graph neural network potential with parallel-MD optimisations (LAMMPS GPU support), enabling very large equivariant simulations while retaining E(3)-equivariance.", "githubUrl": "https://github.com/MDIL-SNU/SevenNet", "paperUrl": "https://arxiv.org/abs/2402.03789", "coverage": [ "general materials" ], "useCases": [ "large-scale MD", "throughput-oriented simulation" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE", "LAMMPS", "PyTorch" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-05", "trainingData": [ "MPTrj" ], "tags": [ "invariant", "speed-optimized", "NequIP-derived" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements covered by MPTrj (~89 elements)", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false, "numElements": 89 }, { "id": "alignn", "type": "node", "category": "Invariant", "label": "ALIGNN", "year": 2021, "author": "NIST", "x": 660, "y": 650, "desc": "Atomistic Line Graph Neural Network: augments the atomic graph with a line graph so bond angles and higher-order interactions are encoded explicitly.", "githubUrl": "https://github.com/usnistgov/alignn", "paperUrl": "https://arxiv.org/abs/2106.01829", "coverage": [ "general materials" ], "useCases": [ "materials property prediction", "JARVIS workflows" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE", "PyTorch" ], "license": "Unlicense", "maintenance": "maintained", "lastReviewed": "2026-05-05", "trainingData": [ "JARVIS-DFT" ], "tags": [ "invariant", "line graph", "directional" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements covered by JARVIS-DFT", "equivariance": "invariant", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "m3gnet", "type": "node", "category": "Invariant", "label": "M3GNet", "year": 2022, "author": "Materials Virtual Lab", "x": 950, "y": 650, "desc": "Materials Graph Network with 3-body interactions; trained on Materials Project relaxations to give a universal potential over most of the periodic table.", "githubUrl": "https://github.com/materialsvirtuallab/m3gnet", "paperUrl": "https://arxiv.org/abs/2202.02450", "coverage": [ "general materials" ], "useCases": [ "universal materials potential", "high-throughput screening" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE" ], "license": "BSD-3-Clause", "maintenance": "archived", "lastReviewed": "2026-05-05", "lastUpdated": "2025-04-09", "trainingData": [ "MPF.2021.2.8" ], "tags": [ "invariant", "foundation model" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "—", "equivariance": "invariant", "architecture": "gnn", "usesAttention": false, "longRange": false, "trainingSetSize": 1800000 }, { "id": "chgnet", "type": "node", "category": "Invariant", "label": "CHGNet", "year": 2023, "author": "Berkeley (Ceder group)", "x": 950, "y": 900, "desc": "Charge-aware graph neural network that extends M3GNet with oxidation state and local charge features; particularly strong for battery and redox-active materials.", "githubUrl": "https://github.com/CederGroupHub/chgnet", "paperUrl": "https://arxiv.org/abs/2302.14231", "coverage": [ "battery materials", "oxides", "redox-active systems" ], "useCases": [ "battery cathode screening", "charge-aware MD" ], "properties": [ "energy", "forces", "stress", "magnetic_moment" ], "frameworks": [ "ASE" ], "license": "BSD-3-Clause", "maintenance": "maintained", "lastReviewed": "2026-05-05", "lastUpdated": "2025-09-22", "trainingData": [ "MPTrj" ], "tags": [ "invariant", "charge-aware", "foundation model" ], "supportsCharges": true, "supportsSpins": true, "elementsCovered": "all elements covered by MPTrj (~89 elements)", "equivariance": "invariant", "architecture": "gnn", "usesAttention": false, "longRange": false, "numElements": 89, "trainingSetSize": 1500000 }, { "id": "sevennet_nano", "type": "node", "category": "Equivariant", "label": "SevenNet-Nano", "year": 2026, "author": "Seoul Nat. Univ. (MDIL-SNU)", "x": 1230, "y": 750, "desc": "Lightweight universal MLIP distilled from the SevenNet-Omni teacher, delivering over an order-of-magnitude speedup while retaining broad transferability for scalable atomistic simulations on thousands of atoms.", "githubUrl": "https://github.com/MDIL-SNU/SevenNet", "paperUrl": "https://arxiv.org/abs/2604.10887", "isNew": true, "coverage": [ "general materials" ], "useCases": [ "lightweight foundation MLIP", "thousand-atom MD" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE", "LAMMPS", "PyTorch" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-05", "trainingData": [ "distilled from SevenNet-Omni" ], "tags": [ "invariant", "distilled", "lightweight", "foundation model" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements covered by SevenNet-Omni", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "sevennet_omni", "type": "node", "category": "Equivariant", "label": "SevenNet-Omni", "year": 2026, "author": "Seoul Nat. Univ. (MDIL-SNU)", "x": 1230, "y": 900, "desc": "Multi-fidelity universal foundation MLIP built on the SevenNet-MF backbone and trained on 15 open datasets (~250M structures across molecules, crystals, and surfaces); serves as the teacher model for SevenNet-Nano.", "githubUrl": "https://github.com/MDIL-SNU/SevenNet", "paperUrl": "https://arxiv.org/abs/2510.11241", "isNew": true, "coverage": [ "general materials", "molecules", "surfaces" ], "useCases": [ "multi-fidelity universal MLIP", "teacher model for distillation" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE", "LAMMPS", "PyTorch" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-05", "trainingData": [ "15 open datasets (~250M structures)" ], "tags": [ "invariant", "multi-fidelity", "foundation model" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements covered by combined open MLIP datasets (~89 elements)", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false, "numElements": 89, "trainingSetSize": 250000000 }, { "id": "pfp_v8", "type": "node", "category": "Transformer", "label": "PFP v8", "year": 2026, "author": "Preferred Networks / Matlantis", "x": 100, "y": 900, "desc": "Eighth release of the Preferred Potential: a universal MLIP trained on a large r2SCAN meta-GGA dataset, capable of reproducing 45 elements off-the-shelf across crystals, molecules, surfaces, and adsorption structures without fine-tuning. Distributed commercially via the Matlantis SaaS platform.", "paperUrl": "https://arxiv.org/abs/2603.11063", "isNew": true, "coverage": [ "general materials" ], "useCases": [ "commercial materials simulation", "Matlantis SaaS" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE" ], "license": "proprietary", "maintenance": "active", "lastReviewed": "2026-05-05", "trainingData": [ "custom r2SCAN", "custom PBE" ], "tags": [ "transformer", "foundation model", "r2SCAN" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements up to Z=96 (PBE backbone) / 70 elements (r2SCAN)", "equivariance": "invariant", "architecture": "gnn", "usesAttention": true, "longRange": false, "numElements": 96 }, { "id": "orion", "type": "node", "category": "Descriptor", "label": "ORION", "year": 2026, "author": "Chen et al. (NEP framework)", "x": 1510, "y": 900, "desc": "Universal organic force field for C, H, O, N, S, P built within the Neuroevolution Potential (NEP) framework. Trained on a chemically rich dataset assembled through a unified top-down/bottom-up sampling strategy, providing a balanced description of bond breaking/formation, aromatic growth, hydrogen bonding, van der Waals interactions, and π-stacking; reaches near-DFT force accuracy while running ~200x faster than ReaxFF on identical hardware, enabling hundreds-of-nanoseconds reactive MD.", "githubUrl": "https://github.com/brucefan1983/GPUMD", "paperUrl": "https://arxiv.org/abs/2604.05769", "isNew": true, "coverage": [ "organic molecules", "reactive chemistry" ], "useCases": [ "reactive MD", "bond breaking/formation", "ReaxFF replacement" ], "properties": [ "energy", "forces" ], "frameworks": [ "LAMMPS" ], "license": "GPL-3.0", "maintenance": "active", "lastReviewed": "2026-05-05", "trainingData": [ "custom organic CHONSP DFT" ], "tags": [ "descriptor", "neuroevolution", "reactive", "organic" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "C, H, O, N, S, P", "equivariance": "invariant", "architecture": "descriptor", "usesAttention": false, "longRange": false, "numElements": 6 }, { "id": "omni_p2x", "type": "node", "category": "Descriptor", "label": "OMNI-P2x", "year": 2026, "author": "Martyka, Tong, Jankowska, Dral (Xiamen University)", "x": 1790, "y": 750, "desc": "First universal neural network potential for molecular ground and excited electronic states. An ensemble of MS-ANI-style invariant potentials trained on PubChemQC TD-DFT (B3LYP/6-31+G*) excited-state data combined with CCSD(T)/CBS ground-state energies from ANI-1ccx, with a separate head predicting oscillator strengths of interstate transitions. Approaches TD-DFT accuracy for UV/vis spectra and photodynamics at a fraction of the cost while outperforming semiempirical methods.", "githubUrl": "https://github.com/dralgroup/omni-p2x", "paperUrl": "https://chemrxiv.org/doi/10.26434/chemrxiv-2025-j207x", "isNew": true, "coverage": [ "organic molecules", "excited states" ], "useCases": [ "UV/vis spectra", "photodynamics" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-05", "properties": [ "energy", "forces" ], "frameworks": [ "ASE" ], "trainingData": [ "PubChemQC", "ANI-1ccx" ], "tags": [ "excited states", "photodynamics", "ensemble" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "H, C, N, O, F, S, Cl", "equivariance": "invariant", "architecture": "descriptor", "usesAttention": false, "longRange": false, "numElements": 7 }, { "id": "grace_off", "type": "node", "category": "Equivariant", "label": "GRACE-OFF", "year": 2026, "author": "Picha, Karwounopoulos, Erhard, Boresch, Heid (TU Wien / U. Vienna)", "x": 1790, "y": 900, "desc": "GRACE-architecture MLIP for organic systems, trained on the SPICE v2.0 dataset and integrated with ASE for MD. Two-layer GRACE-OFF models outperform MACE-OFF (including MACE-OFF24(M)) on single-point energies, forces, torsional profiles, and condensed-phase properties of organic liquids and water; for water and hexane they also beat the much more expensive UMA(S) on densities and radial distribution functions. Established as an accurate, GPU-efficient foundation potential for organic-liquid and biomolecular MD.", "githubUrl": "https://github.com/heid-lab/grace-off", "paperUrl": "https://chemrxiv.org/doi/10.26434/chemrxiv.15001529", "isNew": true, "coverage": [ "organic liquids", "condensed-phase organic chemistry" ], "useCases": [ "organic liquid MD", "RDFs and densities" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-05", "properties": [ "energy", "forces" ], "frameworks": [ "ASE" ], "trainingData": [ "SPICE v2.0" ], "tags": [ "organic liquids", "GRACE", "condensed phase", "reactive" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "SPICE v2.0 organic chemistry coverage (H, C, N, O, F, P, S, Cl, Br, I, plus common ions)", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false, "numElements": 10 }, { "id": "omnimol", "type": "node", "category": "Transformer", "label": "OmniMol", "year": 2026, "author": "Elsharkawy, Mikuni, Bhimji, Nachman (LBNL / NERSC)", "x": 2060, "y": 900, "desc": "Transformer-based small-molecule MLIP that adapts the Omnilearned Point-Edge-Transformer (PET) foundation model — pre-trained on ~1 billion LHC particle jets — to molecular dynamics via cross-domain transfer learning. Uses an interaction-matrix attention bias to inject pairwise atomic physics into transformer attention; on the OMol25 dataset OmniMol-M outperforms a 1B-parameter baseline transformer with ~20× fewer parameters, demonstrating the first cross-discipline transfer for scientific point-cloud foundation models.", "paperUrl": "https://arxiv.org/abs/2601.10791", "isNew": true, "coverage": [ "organic molecules" ], "useCases": [ "molecular MD", "cross-domain foundation transfer" ], "properties": [ "energy", "forces" ], "frameworks": [ "PyTorch" ], "maintenance": "experimental", "lastReviewed": "2026-05-05", "trainingData": [ "OMol25" ], "tags": [ "transformer", "cross-domain transfer", "PET" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "—", "equivariance": "learnt", "architecture": "gnn", "usesAttention": true, "longRange": false }, { "id": "hienet", "type": "node", "category": "Equivariant", "label": "HIENet", "year": 2025, "author": "Yan, Bohde et al. (Texas A&M, divelab/AIRS)", "x": 3180, "y": 320, "desc": "Hybrid Invariant–Equivariant materials foundation potential that interleaves invariant and O(3)-equivariant message-passing layers to leverage invariant-layer scalability while reserving equivariant layers for high-order interactions. Force and stress are obtained as exact derivatives of a conservative energy, and the model achieves SOTA on Matbench Discovery while running ~90% faster than SevenNet-l3i5 and ~140% faster than EquiformerV2.", "githubUrl": "https://github.com/divelab/AIRS", "paperUrl": "https://arxiv.org/abs/2503.05771", "isNew": true, "coverage": [ "general materials" ], "useCases": [ "materials discovery", "Matbench Discovery screening" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE", "PyTorch" ], "license": "GPL-3.0", "maintenance": "active", "lastReviewed": "2026-05-07", "trainingData": [ "MPTrj" ], "tags": [ "equivariant", "hybrid invariant-equivariant", "foundation model" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements covered by MPTrj (~89 elements)", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "mace_mh1", "type": "node", "category": "Equivariant", "label": "MACE-MH-1", "year": 2026, "author": "Batatia, Lin, Hart, Kasoar, Elena, Norwood, Wolf, Csányi (Cambridge / ACEsuit)", "x": 4020, "y": 150, "desc": "Cross-learning multi-head MACE foundation model that bridges molecular, surface, and inorganic crystal chemistry in a single MLIP. Enhances the MACE architecture with stronger element-weight sharing and non-linear tensor-decomposition product bases, then post-trains a multi-head replay scheme on OMAT-24 (PBE crystals), MPTraj, OMol (ωB97M-VV10), OC20 (surfaces), SPICE, RGD1, and MATPES-r2SCAN heads to unify electronic-structure theories.", "githubUrl": "https://github.com/ACEsuit/mace-foundations", "paperUrl": "https://arxiv.org/abs/2510.25380", "isNew": true, "coverage": [ "general materials", "molecules", "surfaces", "molecular crystals" ], "useCases": [ "unified materials/molecular MLIP", "multi-fidelity foundation MD" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE", "LAMMPS", "PyTorch" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-07", "trainingData": [ "OMAT-24", "MPTraj", "OMol", "OC20", "SPICE", "RGD1", "MATPES-r2SCAN" ], "tags": [ "equivariant", "MACE", "multi-head", "cross-domain", "foundation model" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements up to Z=89 (OMAT-24 / MPTraj coverage)", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false, "numElements": 89 }, { "id": "mgnn", "type": "node", "category": "Invariant", "label": "MGNN", "year": 2025, "author": "Chang, Zhu (Wuhan University of Technology)", "x": 2340, "y": 900, "desc": "Moment Graph Neural Network: rotation-invariant message-passing architecture whose node and edge updates operate on Cartesian moment representations of 3D molecular graphs, capturing high-order angular structure without explicit equivariant tensor products. Reaches SOTA on QM9 and revised MD17 (incl. MD17-ethanol) and generalises to 3BPA and 25-element high-entropy alloys, including amorphous-electrolyte MD.", "githubUrl": "https://github.com/JakechiC/MGNN", "paperUrl": "https://arxiv.org/abs/2409.15800", "isNew": true, "coverage": [ "organic molecules", "high-entropy alloys", "amorphous electrolytes" ], "useCases": [ "molecular property prediction", "MD of complex alloys and electrolytes" ], "properties": [ "energy", "forces" ], "frameworks": [ "ASE", "PyTorch" ], "maintenance": "maintained", "lastReviewed": "2026-05-07", "trainingData": [ "QM9", "rMD17", "MD17", "3BPA" ], "tags": [ "invariant", "moment representation", "molecular potential" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "molecular CHNOFS plus 25-element high-entropy alloys", "equivariance": "invariant", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "hydragnn", "type": "node", "category": "Equivariant", "label": "HydraGNN-GFM", "year": 2026, "author": "Lupo Pasini, Choi et al. (ORNL / NERSC)", "x": 2620, "y": 900, "desc": "Exascale multi-task atomistic graph foundation model built on the HydraGNN framework, with a PaiNN-based message-passing backbone selected via large-scale DeepHyper hyperparameter optimization on Frontier. Jointly pre-trained on 16 open first-principles datasets (~544M structures, 85+ elements) using shared message-passing layers and per-dataset output heads, scaled to 16,000 GPUs and able to evaluate 1.1 billion atomistic structures in 50 seconds for downstream materials screening.", "githubUrl": "https://github.com/ORNL/HydraGNN", "paperUrl": "https://arxiv.org/abs/2604.15380", "isNew": true, "coverage": [ "general materials", "organic molecules", "catalysts" ], "useCases": [ "billion-scale screening", "multi-task foundation MLIP" ], "properties": [ "energy", "forces" ], "frameworks": [ "PyTorch", "ASE" ], "license": "BSD-3-Clause", "maintenance": "active", "lastReviewed": "2026-05-08", "trainingData": [ "MPTrj", "OMat24", "OC20", "ANI-1x", "Transition-1x", "OMol25" ], "tags": [ "equivariant", "PaiNN", "multi-task", "exascale", "foundation model" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "85+ elements aggregated across 16 first-principles datasets", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false, "numElements": 85, "trainingSetSize": 544000000 }, { "id": "qnep", "type": "node", "category": "Descriptor", "label": "qNEP", "year": 2026, "author": "Fan, Tang, Berger, Fransson, Xu, Erhart et al. (GPUMD)", "x": 2900, "y": 650, "desc": "Charge-aware extension of the neuroevolution potential (NEP) framework that introduces explicit, environment-dependent partial charges represented per-ion by neural networks of the local descriptor vector. Implemented in GPUMD with both Ewald and particle-particle particle-mesh electrostatics, enabling Born-effective-charge tensors, dielectric properties, and infrared spectra alongside long-range MD scalable to million-atom systems on consumer GPUs.", "githubUrl": "https://github.com/brucefan1983/GPUMD", "paperUrl": "https://arxiv.org/abs/2601.19034", "isNew": true, "coverage": [ "dielectric materials", "ionic systems", "general materials" ], "useCases": [ "dielectric response", "infrared spectra", "long-range MD" ], "properties": [ "energy", "forces", "stress", "dipole", "polarizability" ], "frameworks": [ "LAMMPS" ], "license": "GPL-3.0", "maintenance": "active", "lastReviewed": "2026-05-08", "trainingData": [ "custom DFT" ], "tags": [ "descriptor", "neuroevolution", "charge-aware", "Ewald", "long-range" ], "supportsCharges": true, "supportsSpins": false, "elementsCovered": "dataset-dependent (NEP descriptor framework)", "equivariance": "invariant", "architecture": "descriptor", "usesAttention": false, "longRange": true }, { "id": "petmad15", "type": "node", "category": "Transformer", "label": "PET-MAD-1.5", "year": 2026, "author": "Mazitov, Bigi, Kellner, Pegolo, Tisi, Pozdnyakov, Loche, Kazeev, Fraux, Ceriotti et al. (EPFL / lab-cosmo)", "x": 4020, "y": 320, "desc": "Successor to PET-MAD: a generally applicable r²SCAN universal interatomic potential that extends elemental coverage to 102 elements via the curated MAD-1.5 dataset (~217k structures). Same Point Edge Transformer (PET) backbone with rotation-invariance learnt from data, retrained at the r²SCAN meta-GGA level with targeted enrichment strategies (molecules, clusters, surfaces, low-dimensional structures, bulk crystals) and uncertainty-quantification-driven outlier removal. Reported as more robust, more accurate, and faster than the original PET-MAD across challenging molecular dynamics benchmarks.", "githubUrl": "https://github.com/lab-cosmo/pet-mad", "paperUrl": "https://arxiv.org/abs/2603.02089", "isNew": true, "coverage": [ "general materials", "molecules", "clusters", "surfaces", "low-dimensional structures" ], "useCases": [ "universal MLIP", "r²SCAN molecular dynamics" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE", "LAMMPS" ], "license": "BSD-3-Clause", "maintenance": "active", "lastReviewed": "2026-05-09", "trainingData": [ "MAD-1.5" ], "tags": [ "transformer", "PET", "foundation model", "r²SCAN", "102 elements" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "102 elements", "equivariance": "learnt", "architecture": "gnn", "usesAttention": true, "longRange": false, "numElements": 102, "trainingSetSize": 217000 }, { "id": "densnet", "type": "node", "category": "Equivariant", "label": "DenSNet", "year": 2026, "author": "Lewis-Atwell, Saiz Pardo, Rieu, Loche, Bocus, Cersonsky, Ceriotti et al.", "x": 4300, "y": 150, "desc": "Density-first machine-learned electronic-structure framework that learns the Hohenberg-Kohn map from nuclear configurations to the ground-state electron density using an SE(3)-equivariant neural network predicting density coefficients of an atom-centred Gaussian basis, with a Δ-learning prior built from superposed atomic densities. A second equivariant network then maps the predicted density to the total energy, providing a unified framework for molecular dynamics, energies, forces, and electronic observables (dipole moments, polarizabilities, infrared spectra). Validated on ethanol, ethanethiol, resorcinol, and polythiophene oligomers (extrapolating from 1-6 to 12 monomers).", "paperUrl": "https://arxiv.org/abs/2604.24563", "isNew": true, "coverage": [ "organic molecules", "molecular crystals", "polymers" ], "useCases": [ "IR spectroscopy from MD", "density-aware molecular dynamics", "electronic-structure observables" ], "properties": [ "energy", "forces", "dipole", "polarizability" ], "frameworks": [ "PyTorch" ], "maintenance": "experimental", "lastReviewed": "2026-05-09", "trainingData": [ "custom DFT (ethanol, ethanethiol, resorcinol, polythiophene)" ], "tags": [ "equivariant", "SE(3)", "electron density", "Δ-learning", "spectroscopy" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "—", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "pet_oam_xl", "type": "node", "category": "Transformer", "label": "PET-OAM-XL", "year": 2026, "author": "Bigi, Pegolo, Mazitov, Ceriotti (EPFL / lab-cosmo)", "x": 4300, "y": 320, "desc": "Extra-large variant of the Point Edge Transformer (PET) trained with an OMat24 + sAlex + MPtrj recipe (the OAM data mixture). Pushes the limits of unconstrained MLIPs by trading explicit E(3) symmetry constraints for capacity, depth, and data, achieving the top position on the Matbench Discovery leaderboard at release. Configuration: d_pet 640, d_node 2560, 5 GNN layers + 3 attention layers, 10 Å cutoff with adaptive 40-neighbour cap, distributed via the upet package alongside the wider PET-MAD / PET-OMat / PET-SPICE family.", "githubUrl": "https://github.com/lab-cosmo/upet", "paperUrl": "https://arxiv.org/abs/2601.16195", "isNew": true, "coverage": [ "general materials", "inorganic crystals" ], "useCases": [ "materials discovery", "Matbench Discovery screening", "high-throughput stability prediction" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE", "PyTorch" ], "license": "BSD-3-Clause", "maintenance": "active", "lastReviewed": "2026-05-10", "trainingData": [ "OMat24", "sAlex", "MPtrj" ], "tags": [ "transformer", "PET", "unconstrained", "foundation model", "Matbench Discovery" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements covered by OMat24 / MPtrj (~89 elements)", "equivariance": "learnt", "architecture": "gnn", "usesAttention": true, "longRange": false, "numElements": 89 }, { "id": "alphanet", "type": "node", "category": "Equivariant", "label": "AlphaNet", "year": 2026, "author": "Yin, Zhang, Yang et al.", "x": 4580, "y": 150, "desc": "Local-frame-based equivariant interatomic potential that builds atom-centred frames with learnable geometric transitions, replacing the spherical-harmonic tensor products used by NequIP/MACE-style models with cheaper Cartesian-frame operations. Achieves SOTA accuracy and improved efficiency on Matbench Discovery and OC2M benchmarks across molecular reactions, crystal stability, and surface catalysis, with an OMat24-pretrained foundation variant (alphanet-v1-oam) released on the Matbench Discovery leaderboard.", "githubUrl": "https://github.com/zmyybc/AlphaNet", "paperUrl": "https://arxiv.org/abs/2501.07155", "isNew": true, "coverage": [ "general materials", "organic molecules", "catalysts" ], "useCases": [ "materials discovery", "catalyst screening", "molecular MD" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE", "PyTorch" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-10", "trainingData": [ "OMat24", "MPtrj", "OC2M" ], "tags": [ "equivariant", "local frame", "foundation model" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements covered by OMat24 / MPtrj (~89 elements)", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false, "numElements": 89 }, { "id": "tace", "type": "node", "category": "Equivariant", "label": "TACE", "year": 2026, "author": "Xu et al.", "x": 4580, "y": 320, "desc": "Tensor Atomic Cluster Expansion: a unified Cartesian-space framework that decomposes atomic environments into a complete hierarchy of irreducible Cartesian tensors, providing symmetry-consistent invariant and equivariant representations without spherical-harmonic Clebsch-Gordan overhead. Universal embeddings expose computational level, total charge, magnetic moments, and external-field perturbations as conditioning inputs, while a Latent Ewald Summation module handles long-range electrostatics. Released with an OMat24-pretrained foundation variant (tace-v1-oam-m) on the Matbench Discovery leaderboard, with TorchSim, LAMMPS-ML-IAP, and ASE calculators.", "githubUrl": "https://github.com/xvzemin/tace", "paperUrl": "https://arxiv.org/abs/2509.14961", "isNew": true, "coverage": [ "general materials", "organic molecules", "magnetic systems", "charged systems" ], "useCases": [ "universal MLIP", "tensorial property prediction", "external-field response" ], "properties": [ "energy", "forces", "stress", "dipole", "magnetic_moment", "polarizability" ], "frameworks": [ "ASE", "LAMMPS", "PyTorch" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-10", "trainingData": [ "OMat24", "MPtrj" ], "tags": [ "equivariant", "ACE", "Cartesian tensors", "long-range", "charge-aware", "magnetic", "foundation model" ], "supportsCharges": true, "supportsSpins": true, "elementsCovered": "all elements covered by OMat24 / MPtrj (~89 elements)", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": true, "numElements": 89 }, { "id": "transip", "type": "node", "category": "Transformer", "label": "TransIP", "year": 2026, "author": "Elhag, Raja, Morehead, Blau, Zhao, Tyrchan, Nittinger, Morris, Bronstein (Oxford / LBNL / AstraZeneca / AITHYRA)", "x": 6260, "y": 150, "desc": "Transformer-based interatomic potential built on a generic, non-equivariant backbone that learns SO(3)-equivariance implicitly through a latent equivariance training objective rather than via architectural constraints. Replaces graph-based inductive biases and Clebsch-Gordan tensor products with a scalable embedding-space alignment loss, achieving 40–60% improvement over a data-augmentation baseline across varying OMol25 training-set sizes. Released with a TransIP-L checkpoint on Hugging Face and built on top of the fairchem framework.", "githubUrl": "https://github.com/Ahmed-A-A-Elhag/TransIP", "paperUrl": "https://arxiv.org/abs/2510.00027", "isNew": true, "coverage": [ "organic molecules", "electrolytes", "metal complexes", "biomolecules" ], "useCases": [ "drug discovery", "molecular MD", "scalable MLIP training" ], "properties": [ "energy", "forces" ], "frameworks": [ "ASE", "PyTorch" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-11", "trainingData": [ "OMol25" ], "tags": [ "transformer", "non-equivariant", "latent equivariance", "fairchem" ], "supportsCharges": true, "supportsSpins": true, "elementsCovered": "elements present in OMol25 (organic + electrolyte + metal-complex chemistry)", "equivariance": "learnt", "architecture": "gnn", "usesAttention": true, "longRange": false }, { "id": "e2former", "type": "node", "category": "Transformer", "label": "E2Former", "year": 2025, "author": "Li, Huang, Ding, Wang, Wei, Yang, Wang, Liu, Shi, Jin, Zhang, Gerstein, Qin (Microsoft Research / Yale)", "x": 6540, "y": 150, "desc": "Efficient and equivariant transformer that replaces conventional SO(3) spherical-tensor-product convolutions with a Wigner 6j convolution (Wigner 6j Conv), shifting the dominant compute from edges to nodes and reducing tensor-product complexity from O(|E|) to O(|V|) while preserving E(3) equivariance and expressive power. Achieves a 7×–30× speedup over standard SO(3) convolutions on benchmark interatomic-potential tasks (OC20, OC22, SPICE, MD17/MD22) at chemical accuracy. Released as a NeurIPS 2025 Spotlight and serves as the architectural foundation for the later E2Former-V2 / UBio-MolFM line.", "githubUrl": "https://github.com/liyy2/E2Former", "paperUrl": "https://arxiv.org/abs/2501.19216", "isNew": true, "coverage": [ "catalysts", "surfaces", "organic molecules" ], "useCases": [ "scalable equivariant MLIP", "catalyst screening", "molecular MD" ], "properties": [ "energy", "forces" ], "frameworks": [ "PyTorch" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-14", "trainingData": [ "OC20", "OC22", "SPICE", "MD17", "MD22" ], "tags": [ "transformer", "equivariant", "attention", "Wigner 6j", "linear-scaling" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "dataset-dependent (OC20 / OC22 / SPICE coverage)", "equivariance": "constrained", "architecture": "gnn", "usesAttention": true, "longRange": false }, { "id": "ubio_molfm", "type": "node", "category": "Transformer", "label": "UBio-MolFM", "year": 2026, "author": "Huang, Jiang, Liu, Wang, Zhao, Wang, Lu, Huang, Cheng, Du, Zhang (UBio Team, IQuest Research)", "x": 6820, "y": 150, "desc": "Universal molecular foundation model engineered for biomolecular systems, combining three components: (i) UBio-Mol26, a ~17M-configuration bio-specific dataset built via a multi-fidelity two-pronged strategy that combines systematic bottom-up enumeration with top-down sampling of native protein environments up to ~1,200 atoms (proteins, nucleic acids, lipids, drug-like molecules, biologically important ions and metals); (ii) E2Former-V2, a linear-scaling equivariant transformer integrating Equivariant Axis-Aligned Sparsification (EAAS) and Long-Short Range (LSR) modelling for ~4× higher inference throughput on large systems; and (iii) a three-stage curriculum-learning protocol going from energy initialisation to energy-force consistency and multi-fidelity refinement. Targets ab-initio-level fidelity on out-of-distribution biomolecular systems up to ~1,500 atoms.", "githubUrl": "https://github.com/IQuestLab/UBio-MolFM", "paperUrl": "https://arxiv.org/abs/2602.17709", "isNew": true, "coverage": [ "biomolecules", "proteins", "nucleic acids", "lipids", "drug-like molecules", "explicit solvent" ], "useCases": [ "biomolecular MD", "DFT-fidelity bio-system simulation", "drug discovery" ], "properties": [ "energy", "forces" ], "frameworks": [ "PyTorch" ], "maintenance": "experimental", "lastReviewed": "2026-05-14", "trainingData": [ "UBio-Mol26" ], "tags": [ "transformer", "equivariant", "attention", "bio-systems", "foundation model", "E2Former-V2" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "biological organic chemistry (CHNOPS + halogens) plus biologically important ions and metals", "equivariance": "constrained", "architecture": "gnn", "usesAttention": true, "longRange": true }, { "id": "fennix_bio1", "type": "node", "category": "Equivariant", "label": "FeNNix-Bio1", "year": 2026, "author": "Plé, Adjoua, Benali, Posenitskiy, Villot, Lagardère, Piquemal (Sorbonne Université / Qubit Pharmaceuticals)", "x": 6260, "y": 320, "desc": "Force-field-enhanced foundation MLIP for drug design built on the FENNIX hybrid framework: an Allegro E(3)-equivariant local embedding (256 scalar features, 16 equivariant channels up to l_max=2, three interaction layers, 5.3 Å cutoff) coupled with physics-grounded energy terms — short-range ZBL repulsion, Coulomb electrostatics with fluctuating atom-centred charges, and explicit dispersion. Trained on an extended SPICE2 synthetic-QM dataset and integrated with Tinker-HP for reactive molecular dynamics including quantum nuclear effects. Sets a new sub-kcal/mol standard on the Freesolv hydration-free-energy benchmark and supports protein folding, protein-ligand binding free energies, and chemical reactions.", "githubUrl": "https://github.com/thomasple/FeNNol", "paperUrl": "https://doi.org/10.26434/chemrxiv-2025-f1hgn", "isNew": true, "coverage": [ "organic molecules", "biomolecules", "proteins", "electrolytes" ], "useCases": [ "drug discovery", "hydration free energies", "protein folding", "protein-ligand binding", "reactive MD" ], "properties": [ "energy", "forces" ], "frameworks": [ "ASE", "PyTorch" ], "license": "LGPL-3.0", "maintenance": "active", "lastReviewed": "2026-05-15", "trainingData": [ "SPICE2", "extended SPICE2" ], "tags": [ "equivariant", "force-field-enhanced", "biomolecular", "drug design", "foundation model", "JAX", "charge-aware" ], "supportsCharges": true, "supportsSpins": false, "elementsCovered": "biomolecular elements (H, C, N, O, F, P, S, Cl, Br, I) plus Li, Na, K, Mg, Zn", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": true }, { "id": "dtnn", "type": "node", "category": "Invariant", "label": "DTNN", "year": 2017, "author": "Schütt, Arbabzadah, Chmiela, Müller, Tkatchenko", "x": 2900, "y": 550, "desc": "Deep Tensor Neural Network — the direct precursor to SchNet that introduced learned per-element embeddings refined by tensorised pairwise interaction blocks for quantum-chemical energies on QM9-style organic molecules.", "githubUrl": "https://github.com/atomistic-machine-learning/dtnn", "paperUrl": "https://www.nature.com/articles/ncomms13890", "coverage": [ "organic molecules" ], "useCases": [ "molecular property prediction", "GNN MLIP precursor" ], "properties": [ "energy" ], "frameworks": [ "PyTorch" ], "license": "MIT", "maintenance": "archived", "lastReviewed": "2026-05-10", "trainingData": [ "QM9", "GDB-9" ], "tags": [ "invariant", "tensor", "deep network", "precursor to SchNet" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "dataset-dependent (organic molecules)", "equivariance": "invariant", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "cgcnn", "type": "node", "category": "Invariant", "label": "CGCNN", "year": 2018, "author": "Xie & Grossman (MIT)", "x": 2900, "y": 900, "desc": "Crystal Graph Convolutional Neural Network — the first GNN built explicitly on periodic crystal graphs with multi-edge bond convolutions. Foundational ancestor of nearly every later universal materials GNN (MEGNet / M3GNet / CHGNet) and a workhorse for materials property prediction.", "githubUrl": "https://github.com/txie-93/cgcnn", "paperUrl": "https://arxiv.org/abs/1710.10324", "coverage": [ "periodic crystals", "general materials" ], "useCases": [ "materials property prediction", "crystal-graph GNN baseline" ], "properties": [ "energy" ], "frameworks": [ "PyTorch" ], "license": "MIT", "maintenance": "archived", "lastReviewed": "2026-05-10", "trainingData": [ "Materials Project" ], "tags": [ "invariant", "crystal graph", "convolutional", "foundation" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements covered by Materials Project", "equivariance": "invariant", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "physnet", "type": "node", "category": "Invariant", "label": "PhysNet", "year": 2019, "author": "Unke & Meuwly (Basel)", "x": 3180, "y": 550, "desc": "Modular invariant message-passing potential that simultaneously predicts energies, forces, dipole moments, and partial charges with explicit electrostatic and dispersion energy terms. One of the first MLIPs that handled molecules with non-zero net charge through learnable atomic charges plus Coulomb correction.", "githubUrl": "https://github.com/MMunibas/PhysNet", "paperUrl": "https://arxiv.org/abs/1902.08408", "coverage": [ "organic molecules", "small-molecule reactions" ], "useCases": [ "molecular dynamics", "charge-aware MLIP", "dipole prediction" ], "properties": [ "energy", "forces", "dipole" ], "frameworks": [ "ASE" ], "license": "MIT", "maintenance": "archived", "lastReviewed": "2026-05-10", "trainingData": [ "QM9", "MD17", "ANI-1x" ], "tags": [ "invariant", "message-passing", "charge-aware", "electrostatics" ], "supportsCharges": true, "supportsSpins": false, "elementsCovered": "dataset-dependent (organic molecules)", "equivariance": "invariant", "architecture": "gnn", "usesAttention": false, "longRange": true }, { "id": "megnet", "type": "node", "category": "Invariant", "label": "MEGNet", "year": 2019, "author": "Chen, Ye, Zuo, Zheng, Ong (UCSD / Materials Virtual Lab)", "x": 3180, "y": 650, "desc": "MatErials Graph Network — universal graph network with global state attributes that unifies molecules and crystals in one framework. Direct architectural ancestor of M3GNet/CHGNet from the same group; widely used for materials property prediction.", "githubUrl": "https://github.com/materialsvirtuallab/megnet", "paperUrl": "https://pubs.acs.org/doi/10.1021/acs.chemmater.9b01294", "coverage": [ "organic molecules", "periodic crystals", "general materials" ], "useCases": [ "materials property prediction", "molecule-crystal joint modeling" ], "properties": [ "energy" ], "frameworks": [ "PyTorch", "ASE" ], "license": "BSD-3-Clause", "maintenance": "maintained", "lastReviewed": "2026-05-10", "trainingData": [ "Materials Project", "QM9" ], "tags": [ "invariant", "graph network", "global attributes", "materials" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements covered by Materials Project", "equivariance": "invariant", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "cormorant", "type": "node", "category": "Equivariant", "label": "Cormorant", "year": 2019, "author": "Anderson, Hy, Kondor (Chicago)", "x": 4860, "y": 320, "desc": "Covariant Molecular Neural Network — an early end-to-end SO(3)-equivariant architecture using Clebsch-Gordan tensor products on irreducible representations of irreducible spherical tensors. Predates and informs the TFN/NequIP-style equivariant message-passing potential family.", "githubUrl": "https://github.com/risilab/cormorant", "paperUrl": "https://arxiv.org/abs/1906.04015", "coverage": [ "small molecules" ], "useCases": [ "equivariance research baseline", "molecular property prediction" ], "properties": [ "energy" ], "frameworks": [ "PyTorch" ], "license": "Custom (research)", "maintenance": "archived", "lastReviewed": "2026-05-10", "trainingData": [ "QM9", "MD17" ], "tags": [ "equivariant", "SO(3)", "Clebsch-Gordan", "tensor product" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "dataset-dependent (organic molecules)", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "egnn", "type": "node", "category": "Equivariant", "label": "EGNN", "year": 2021, "author": "Satorras, Hoogeboom, Welling (UvA)", "x": 4860, "y": 150, "desc": "E(n)-Equivariant Graph Neural Network: a simple, scalar-only equivariant architecture that achieves rotation/translation equivariance without higher-order tensor representations. Highly cited for its simplicity and broadly applied beyond MLIPs to generative modelling, protein structure, and dynamics.", "githubUrl": "https://github.com/vgsatorras/egnn", "paperUrl": "https://arxiv.org/abs/2102.09844", "coverage": [ "small molecules", "general point clouds" ], "useCases": [ "equivariant baseline", "molecular dynamics", "generative modeling" ], "properties": [ "energy", "forces" ], "frameworks": [ "PyTorch" ], "license": "MIT", "maintenance": "archived", "lastReviewed": "2026-05-10", "trainingData": [ "QM9", "N-body" ], "tags": [ "equivariant", "scalar", "lightweight", "E(n)" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "dataset-dependent", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "nep_orig", "type": "node", "category": "Descriptor", "label": "NEP", "year": 2021, "author": "Fan, Wang, Ying et al. (Aalto / GPUMD)", "x": 3180, "y": 750, "desc": "Original Neuroevolution Potential — a per-element neural network on local descriptors trained with separable natural-evolution strategies rather than gradient descent. Designed for raw GPU throughput in the GPUMD code; ancestor of NEP89, qNEP, and ORION.", "githubUrl": "https://github.com/brucefan1983/GPUMD", "paperUrl": "https://arxiv.org/abs/2107.08119", "coverage": [ "general materials", "small molecules" ], "useCases": [ "GPU MD throughput", "ML potential framework" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "LAMMPS" ], "license": "GPL-3.0", "maintenance": "active", "lastReviewed": "2026-05-10", "trainingData": [ "custom DFT" ], "tags": [ "descriptor", "neuroevolution", "GPU MD", "GPUMD" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "dataset-dependent", "equivariance": "invariant", "architecture": "descriptor", "usesAttention": false, "longRange": false }, { "id": "pfp_orig", "type": "node", "category": "Invariant", "label": "PFP", "year": 2022, "author": "Takamoto et al. (Preferred Networks)", "x": 3180, "y": 900, "desc": "Original Preferred Potential — first universal NNP covering 45 elements and the foundation of the commercial Matlantis SaaS platform. Predated the academic universal MLIP wave (M3GNet, CHGNet, MACE-MP) by over a year and demonstrated viable industrial-scale deployment.", "paperUrl": "https://www.nature.com/articles/s41467-022-30687-9", "coverage": [ "general materials", "catalysts", "surfaces" ], "useCases": [ "commercial materials simulation", "first universal NNP" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE" ], "license": "proprietary", "maintenance": "maintained", "lastReviewed": "2026-05-10", "trainingData": [ "custom DFT" ], "tags": [ "invariant", "TeaNet", "foundation model", "Matlantis" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "45 elements (Matlantis PFP-1.0)", "numElements": 45, "equivariance": "invariant", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "equiformer_v1", "type": "node", "category": "Transformer", "label": "Equiformer", "year": 2022, "author": "Liao & Smidt (MIT Atomic Architects)", "x": 5140, "y": 150, "desc": "Original Equivariant Graph Attention Transformer — combines graph attention with TFN-style E(3)-equivariant tensor representations and depthwise tensor-product MLPs. Direct ancestor of Equiformer V2 and V3; influential for attention-based equivariant MLIPs.", "githubUrl": "https://github.com/atomicarchitects/equiformer", "paperUrl": "https://arxiv.org/abs/2206.11990", "coverage": [ "organic molecules", "catalysts", "general materials" ], "useCases": [ "equivariant transformer", "OC20 baseline" ], "properties": [ "energy", "forces" ], "frameworks": [ "PyTorch", "ASE" ], "license": "MIT", "maintenance": "archived", "lastReviewed": "2026-05-10", "trainingData": [ "QM9", "MD17", "OC20" ], "tags": [ "transformer", "equivariant", "attention", "TFN-style" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "dataset-dependent", "equivariance": "constrained", "architecture": "gnn", "usesAttention": true, "longRange": false }, { "id": "scn", "type": "node", "category": "Transformer", "label": "SCN", "year": 2022, "author": "Zitnick, Das, Goyal et al. (Meta FAIR)", "x": 5140, "y": 320, "desc": "Spherical Channels Network — represents atomic environments as multichannel spherical signals rotated into edge-aligned local frames, enabling efficient high-degree representations for OC20-class catalyst modelling. Direct precursor to eSCN and the Meta FAIR equivariant-transformer line.", "githubUrl": "https://github.com/Open-Catalyst-Project/ocp", "paperUrl": "https://arxiv.org/abs/2206.14331", "coverage": [ "catalysts", "surfaces", "general materials" ], "useCases": [ "catalyst discovery", "OC20 modeling" ], "properties": [ "energy", "forces" ], "frameworks": [ "PyTorch" ], "license": "MIT", "maintenance": "archived", "lastReviewed": "2026-05-10", "trainingData": [ "OC20" ], "tags": [ "equivariant", "spherical channels", "edge frames" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements covered by OC20 (~56 elements)", "numElements": 56, "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "so3krates", "type": "node", "category": "Equivariant", "label": "SO3krates", "year": 2022, "author": "Frank, Unke, Müller (TU Berlin)", "x": 5420, "y": 150, "desc": "SO(3)-equivariant attention on arbitrary length scales: factorises equivariant tensor products into invariant scalar attention plus an equivariant filter, enabling long-range-capable equivariant transformers. Foundation of the later SO3LR molecular response model.", "githubUrl": "https://github.com/thorben-frank/mlff", "paperUrl": "https://arxiv.org/abs/2205.14276", "coverage": [ "organic molecules", "biomolecules" ], "useCases": [ "long-range equivariant MD", "spectroscopy" ], "properties": [ "energy", "forces" ], "frameworks": [ "JAX-MD" ], "license": "MIT", "maintenance": "maintained", "lastReviewed": "2026-05-10", "trainingData": [ "QM7-X", "MD17" ], "tags": [ "equivariant", "attention", "long-range", "SO(3)" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "dataset-dependent (organic molecules)", "equivariance": "constrained", "architecture": "gnn", "usesAttention": true, "longRange": true }, { "id": "visnet", "type": "node", "category": "Equivariant", "label": "ViSNet", "year": 2022, "author": "Wang, Zhao, Cui et al. (Microsoft Research)", "x": 5420, "y": 320, "desc": "Vector-scalar interactive message passing potential that captures geometric information through coupled scalar and vector channels without explicit higher-order tensor algebra. Combines competitive accuracy on QM9/MD17/MD22 with practical efficiency for biomolecular MD.", "githubUrl": "https://github.com/microsoft/AI2BMD/tree/ViSNet", "paperUrl": "https://arxiv.org/abs/2210.16518", "coverage": [ "organic molecules", "biomolecules" ], "useCases": [ "biomolecular MD", "molecular property prediction" ], "properties": [ "energy", "forces" ], "frameworks": [ "PyTorch" ], "license": "MIT", "maintenance": "maintained", "lastReviewed": "2026-05-10", "trainingData": [ "QM9", "MD17", "MD22", "Chignolin" ], "tags": [ "equivariant", "vector-scalar", "biomolecular" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "dataset-dependent", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "dpa1", "type": "node", "category": "Descriptor", "label": "DPA-1", "year": 2022, "author": "Zhang, Bi et al. (DeepModeling)", "x": 3460, "y": 550, "desc": "First-generation Deep Potential with Attention — pretrained Deep Potential descriptor model that introduces an attention layer to the DeepMD framework, enabling cross-domain transfer learning. Direct bridge between DeepMD and DPA-2 in the DeepModeling lineage.", "githubUrl": "https://github.com/deepmodeling/deepmd-kit", "paperUrl": "https://arxiv.org/abs/2208.08236", "coverage": [ "general materials", "alloys" ], "useCases": [ "pretrained descriptor", "transfer learning" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "PyTorch", "LAMMPS" ], "license": "LGPL-3.0", "maintenance": "maintained", "lastReviewed": "2026-05-10", "trainingData": [ "custom DFT", "OC20" ], "tags": [ "descriptor", "attention", "pretrained", "DeepMD" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "dataset-dependent", "equivariance": "invariant", "architecture": "descriptor", "usesAttention": true, "longRange": false }, { "id": "escn", "type": "node", "category": "Transformer", "label": "eSCN", "year": 2023, "author": "Passaro & Zitnick (Meta FAIR)", "x": 5700, "y": 150, "desc": "Reduces SO(3) tensor-product convolutions to SO(2) by aligning each pair to a common rotation axis, dramatically lowering the cost of high-degree equivariant convolutions. Direct precursor to Equiformer V2's higher-degree backbone and key SOTA on OC20 in 2023.", "githubUrl": "https://github.com/Open-Catalyst-Project/ocp", "paperUrl": "https://arxiv.org/abs/2302.03655", "coverage": [ "catalysts", "surfaces", "general materials" ], "useCases": [ "catalyst discovery", "high-degree equivariant convolutions" ], "properties": [ "energy", "forces" ], "frameworks": [ "PyTorch" ], "license": "MIT", "maintenance": "maintained", "lastReviewed": "2026-05-10", "trainingData": [ "OC20" ], "tags": [ "equivariant", "SO(3)→SO(2)", "high-degree" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements covered by OC20 (~56 elements)", "numElements": 56, "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "mace_off", "type": "node", "category": "Equivariant", "label": "MACE-OFF", "year": 2023, "author": "Kovács, Moore, Browning, Batatia, Horton et al.", "x": 5980, "y": 150, "desc": "Transferable MACE force field for organic molecules covering 10 elements (H, C, N, O, F, P, S, Cl, Br, I) trained on SPICE quantum-chemistry data. The molecular sibling of MACE-MP-0 and template for later GRACE-OFF; widely used as a drop-in replacement for classical biomolecular force fields.", "githubUrl": "https://github.com/ACEsuit/mace-off", "paperUrl": "https://arxiv.org/abs/2312.15211", "coverage": [ "organic molecules", "biomolecules", "drug-like molecules" ], "useCases": [ "organic MD", "biomolecular simulation", "drug discovery" ], "properties": [ "energy", "forces", "dipole" ], "frameworks": [ "ASE", "LAMMPS" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-10", "trainingData": [ "SPICE" ], "tags": [ "equivariant", "organic FF", "MACE family", "biomolecular" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "H, C, N, O, F, P, S, Cl, Br, I", "numElements": 10, "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "gnome", "type": "node", "category": "Invariant", "label": "GNoME", "year": 2023, "author": "Merchant, Batzner, Schoenholz, Aykol, Cheon, Cubuk (Google DeepMind)", "x": 3460, "y": 650, "desc": "Graph Networks for Materials Exploration — published in Nature 624, 80 (2023), demonstrated discovery of 2.2M new crystal structures (380k stable) via active-learning-coupled NNP-driven materials search. Established graph-network MLIPs as a viable engine for autonomous crystal discovery at scale.", "paperUrl": "https://www.nature.com/articles/s41586-023-06735-9", "githubUrl": "https://github.com/google-deepmind/materials_discovery", "coverage": [ "periodic crystals", "general materials" ], "useCases": [ "materials discovery", "stable structure prediction" ], "properties": [ "energy", "forces" ], "frameworks": [ "JAX-MD" ], "license": "Apache-2.0", "maintenance": "maintained", "lastReviewed": "2026-05-10", "trainingData": [ "custom DFT", "Materials Project" ], "tags": [ "invariant", "graph network", "materials discovery", "active learning" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements covered by Materials Project", "equivariance": "invariant", "architecture": "gnn", "usesAttention": false, "longRange": false, "numElements": 89 }, { "id": "mace_mp0", "type": "node", "category": "Equivariant", "label": "MACE-MP-0", "year": 2024, "author": "Batatia, Benner, Chiang, Elena, Kovács et al.", "x": 5700, "y": 320, "desc": "First MACE foundation model — a single MACE-architecture potential trained on the Materials Project trajectory dataset (MPtrj) covering 89 elements. Demonstrated broadly transferable accuracy across inorganic crystals, surfaces, defects, and molecular crystals; the reference universal MACE model and ancestor of MACE-Osaka26 / MACE-MH-1 / MACE-Magnetic / MACE-POLAR-1.", "githubUrl": "https://github.com/ACEsuit/mace-foundations", "paperUrl": "https://arxiv.org/abs/2401.00096", "coverage": [ "general materials", "organic molecules", "oxides", "molecular crystals" ], "useCases": [ "universal MLIP", "transferable foundation model", "out-of-the-box MD" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE", "LAMMPS" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-10", "trainingData": [ "MPtrj" ], "tags": [ "equivariant", "foundation model", "universal", "MACE family" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements covered by MPtrj (~89 elements)", "numElements": 89, "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false }, { "id": "mpnice", "type": "node", "category": "Invariant", "label": "MPNICE", "year": 2025, "author": "Weber et al. (Schrödinger Inc. / Columbia University)", "x": 4020, "y": 750, "desc": "Message Passing Network with Iterative Charge Equilibration: an invariant message-passing MLFF that decomposes the total energy into atom-centred contributions and, after every message-passing block, predicts atomic partial charges via a charge-equilibration (Qeq) approximation that feeds back into the next block. Explicit long-range Coulomb electrostatics on top of the equilibrated charges captures polarisation, charge transfer, and multiple oxidation states while remaining 5-20x faster than comparable-accuracy models. Released with pre-trained MPNICE models spanning 89 elements for organic liquids, molecular crystals, and bulk inorganic systems.", "paperUrl": "https://arxiv.org/abs/2505.06462", "isNew": true, "coverage": [ "organic liquids", "molecular crystals", "bulk inorganic" ], "useCases": [ "liquid properties", "structural ranking", "ionic / charge-state simulations" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE" ], "license": "proprietary", "maintenance": "active", "lastReviewed": "2026-05-13", "trainingData": [ "custom organic + inorganic DFT" ], "tags": [ "invariant", "charge-aware", "iterative charge equilibration", "long-range-electrostatics" ], "supportsCharges": true, "supportsSpins": false, "elementsCovered": "all elements up to Z=94 (89-element pretrained set)", "equivariance": "invariant", "architecture": "gnn", "usesAttention": false, "longRange": true, "numElements": 89 }, { "id": "omni_p1", "type": "node", "category": "Descriptor", "label": "OMNI-P1", "year": 2025, "author": "Chen & Dral (Xiamen University)", "x": 4300, "y": 750, "desc": "First-ever universal interatomic potential that simultaneously learns and predicts at multiple quantum-chemical levels of theory through a multi-fidelity AIO-ANI training strategy. Combines low- and high-level reference data (semiempirical, DFT, and coupled cluster) into a single network so the same model can deliver predictions across QC levels of organic chemistry with accuracy comparable to GFN2-xTB and double-zeta DFT, while being orders of magnitude faster. Released as part of the MLatom ecosystem with model weights and code in the dralgroup/aio-ani repository; the immediate predecessor to OMNI-P2x for excited states.", "githubUrl": "https://github.com/dralgroup/aio-ani", "paperUrl": "https://doi.org/10.1021/acs.jctc.5c00858", "isNew": true, "coverage": [ "organic molecules" ], "useCases": [ "multi-fidelity geometry optimisation", "DFT replacement", "general organic chemistry" ], "properties": [ "energy", "forces" ], "frameworks": [ "ASE" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-13", "trainingData": [ "ANI-1ccx", "ANI-1x", "GFN2-xTB reference data" ], "tags": [ "multi-fidelity", "universal", "ANI family", "AIO-ANI" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "H, C, N, O", "equivariance": "invariant", "architecture": "descriptor", "usesAttention": false, "longRange": false, "numElements": 4 }, { "id": "geodite", "type": "node", "category": "Equivariant", "label": "Geodite", "year": 2026, "author": "Reschützegger, Aykent, Perin, Nunes, Cipcigan, Ferreira, Steiner et al. (IBM Research / U. São Paulo / Politecnico Milano)", "x": 5980, "y": 320, "desc": "Equivariant message-passing interatomic potential that replaces Clebsch-Gordan tensor products with cheaper geometric operations while keeping E(3) equivariance, and bakes in physical priors (ZBL short-range repulsion, smooth attenuation) so the learned potential energy surface stays smooth and well-behaved across the entire bond-length range. The Geodite-MP variant is trained on the Materials Project trajectory dataset (~160k inorganic crystal structures) and matches the accuracy of leading tensor-product foundation potentials on stability prediction, phonon-derived properties, thermal conductivity, and nanosecond MD, while running 3-5x faster than models with comparable accuracy.", "paperUrl": "https://arxiv.org/abs/2601.15492", "isNew": true, "coverage": [ "general materials", "inorganic crystals" ], "useCases": [ "materials stability prediction", "phonons", "thermal conductivity", "large-scale MD" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-13", "trainingData": [ "MPtrj" ], "tags": [ "equivariant", "no tensor products", "smooth PES", "ZBL", "foundation model" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "all elements covered by MPtrj (~89 elements)", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false, "numElements": 89 }, { "id": "lorem", "type": "node", "category": "Equivariant", "label": "LOREM", "year": 2026, "author": "Rumiantsev, Langer, Sodjargal, Ceriotti, Loche (EPFL lab-cosmo)", "x": 6540, "y": 320, "desc": "Equivariant message-passing interatomic potential built around equivariant — rather than scalar — charges for long-range interactions. Unlike standard cutoff-based MLIPs that fail to capture electrostatics, dispersion, or electron delocalisation, LOREM propagates equivariant charges through a long-range message-passing mechanism that captures orientation-dependent interactions without requiring per-dataset tuning of interaction cutoffs or message-passing depth. Implemented in JAX/Flax on top of e3x and jax-pme, LOREM is competitive with or surpasses prior scalar-charge long-range MLIPs across long-range benchmark datasets. Accepted at TMLR.", "githubUrl": "https://github.com/lab-cosmo/lorem-jax", "paperUrl": "https://arxiv.org/abs/2507.19382", "isNew": true, "coverage": [ "general materials", "organic molecules", "long-range systems" ], "useCases": [ "long-range MLIP", "electrostatics-aware MD", "dispersion-aware MD" ], "properties": [ "energy", "forces" ], "frameworks": [ "ASE", "JAX-MD" ], "license": "BSD-3-Clause", "maintenance": "active", "lastReviewed": "2026-05-17", "trainingData": [ "long-range benchmark datasets" ], "tags": [ "equivariant", "long-range", "equivariant charges", "JAX", "e3x", "jax-pme" ], "supportsCharges": true, "supportsSpins": false, "elementsCovered": "—", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": true, "numElements": null }, { "id": "nemp", "type": "node", "category": "Equivariant", "label": "NEMP", "year": 2026, "author": "Zhang & Guo (University of New Mexico)", "x": 6820, "y": 320, "desc": "Node-equivariant message-passing graph neural network that reformulates equivariant message passing from edges to nodes: rather than computing a tensor product between every edge and its neighbour node, NEMP constructs an expressive equivariant structure via a single tensor product between the central node and a virtual summed node encoding its neighbours. This reduces the cost of the equivariant message-passing step from the number of neighbour pairs to the number of central atoms, achieving 1–2 orders of magnitude reduction in memory and compute over edge-equivariant baselines while matching or exceeding accuracy across molecules, extended systems, and universal-potential benchmarks. Published in Chemical Science (RSC, 2026).", "githubUrl": "https://github.com/zhangylch/NEMP", "paperUrl": "https://arxiv.org/abs/2508.16086", "isNew": true, "coverage": [ "organic molecules", "extended systems", "general materials" ], "useCases": [ "large-scale equivariant MD", "efficient equivariant MLIP" ], "properties": [ "energy", "forces" ], "frameworks": [ "ASE", "JAX-MD" ], "maintenance": "active", "lastReviewed": "2026-05-17", "trainingData": [ "MD17", "MD22" ], "tags": [ "equivariant", "node-equivariant", "efficient message passing", "JAX" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "—", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false, "numElements": null }, { "id": "mace4ir", "type": "node", "category": "Equivariant", "label": "MACE4IR", "year": 2025, "author": "Bhatia, Krejci, Botti, Rinke, Marques (Aalto / Friedrich Schiller University Jena / Ruhr University Bochum)", "x": 7100, "y": 320, "desc": "Foundation model for molecular infrared spectroscopy built on the MACE E(3)-equivariant message-passing architecture. The model consists of two parallel MACE networks — one for energies and forces, one for atomic dipole moments — trained on ~16 million molecular geometries with DFT energies, forces, and dipole moments from the QCML dataset (≈80 elements, organic molecules, inorganic species, and metal complexes). The MACE4IRmol variant (v2) wraps an ensemble of these networks for uncertainty quantification and delivers nuclear-quantum-effects-aware infrared spectra with DFT-level accuracy at a fraction of the cost. Released with model weights on Hugging Face and the QCML-derived AIMD trajectories on Zenodo.", "paperUrl": "https://arxiv.org/abs/2508.19118", "isNew": true, "coverage": [ "organic molecules", "inorganic species", "metal complexes" ], "useCases": [ "infrared spectra", "dipole moment prediction", "molecular MD", "uncertainty quantification" ], "properties": [ "energy", "forces", "dipole" ], "frameworks": [ "ASE", "PyTorch" ], "maintenance": "active", "lastReviewed": "2026-05-18", "trainingData": [ "QCML" ], "tags": [ "equivariant", "MACE family", "IR spectroscopy", "dipole moments", "foundation model", "uncertainty quantification" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "~80 elements (QCML coverage)", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": false, "numElements": 80 }, { "id": "mattersim_mt", "type": "node", "category": "Equivariant", "label": "MatterSim-MT", "year": 2026, "author": "Yang, Liu, Hu, Zhou, Shi, C. Liu, Tan, Li, Wang, Zhu, Chen, Thiemann, Zeni, Horton, Pinsler, Fowler, Zügner, Xie, Sun, Chen, Kong, Bai, Gunceler, Noé, Hao, Lu et al. (Microsoft Research)", "x": 7100, "y": 150, "desc": "Multi-task extension of the MatterSim foundation MLIP that predicts energies, forces, and stress jointly with electronic and tensorial materials properties — Bader charges, magnetic moments, Born effective charges, and full dielectric matrices — from a single shared atomistic backbone. Pretrained on 35M+ first-principles-labelled structures spanning 89 elements, temperatures up to 5000 K, and pressures up to 1000 GPa, then fine-tuned on the property-specific heads. Demonstrated on vibrational spectroscopy, ferroelectric switching, and electrochemical redox processes — workflows that conventional potential-energy-surface MLIPs cannot resolve out of the box.", "githubUrl": "https://github.com/microsoft/mattersim", "paperUrl": "https://arxiv.org/abs/2605.07927", "isNew": true, "coverage": [ "general materials", "ferroelectrics", "dielectrics", "magnetic materials" ], "useCases": [ "multi-task property prediction", "vibrational spectroscopy", "ferroelectric switching", "electrochemical redox" ], "properties": [ "energy", "forces", "stress", "magnetic_moment", "polarizability" ], "frameworks": [ "ASE", "PyTorch" ], "license": "MIT", "maintenance": "active", "lastReviewed": "2026-05-19", "trainingData": [ "custom MatterSim DFT" ], "tags": [ "transformer", "foundation model", "multi-task", "Born effective charges", "Bader charges", "dielectric tensor" ], "supportsCharges": true, "supportsSpins": true, "elementsCovered": "all elements up to Z=89", "equivariance": "constrained", "architecture": "gnn", "usesAttention": true, "longRange": false, "numElements": 89 }, { "id": "polymlp", "type": "node", "category": "Descriptor", "label": "PolyMLP", "year": 2023, "author": "Seko (Kyoto University)", "x": 3740, "y": 550, "desc": "Polynomial machine learning potential representing the atomic energy as a polynomial function of linearly independent polynomial invariants of the O(3) group built from local structural features. A systematically improvable linear/polynomial descriptor MLIP in the SNAP/MTP/ACE family, distributed via the pypolymlp toolkit and the curated Polynomial MLP Repository, and shipped to OpenKIM in March 2026 as a portable model driver with 837 elemental and alloy parameterizations for LAMMPS-scale simulations.", "githubUrl": "https://github.com/sekocha/pypolymlp", "paperUrl": "https://doi.org/10.1063/5.0129045", "isNew": true, "coverage": [ "metals", "alloys", "general materials" ], "useCases": [ "industrial-scale linear MLIP", "high-throughput MD", "phonon calculations", "alloy systems" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "LAMMPS", "ASE", "OpenKIM" ], "license": "BSD-3-Clause", "maintenance": "active", "lastReviewed": "2026-05-20", "trainingData": [ "custom DFT" ], "tags": [ "descriptor", "linear", "polynomial", "O(3) invariants", "pypolymlp", "OpenKIM" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "broad elemental and alloy coverage via the OpenKIM PolyMLP model driver (837 parameterizations)", "equivariance": "invariant", "architecture": "descriptor", "usesAttention": false, "longRange": false }, { "id": "acnn", "type": "node", "category": "Descriptor", "label": "ACNN", "year": 2026, "author": "Li, Feng, Luo, Jiang, Zheng, Song, Lv, Butler, Liu, Xie, Xie, Ma (Jilin University / CALYPSO)", "x": 4580, "y": 550, "desc": "Attention-Coupled Neural Network potential: an expert MLIP architecture built from four modules — a local descriptor, an elemental embedding, a stack of optional multi-head attention modules, and an MLP-based fitting head. Stacking attention modules implicitly introduces an aggregative effect analogous to graph neural networks, extending the effective interaction range among atoms and enabling adaptive switching between network configurations based on the simulation task. Embedded in a self-optimizing CALYPSO crystal-structure-prediction workflow that iteratively refines the potential while exploring local minima of the potential energy surface, validated on Mg-Ca-H ternary and Be-P-N-O quaternary systems by exploring nearly 10 million configurations with substantial speedup over first-principles calculations.", "paperUrl": "https://doi.org/10.1038/s41524-026-01971-9", "isNew": true, "coverage": [ "alloys", "ternary systems", "quaternary systems", "high-pressure hydrides" ], "useCases": [ "crystal structure prediction", "high-throughput structural optimization", "automated material discovery" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE" ], "maintenance": "active", "lastReviewed": "2026-05-22", "trainingData": [ "custom DFT" ], "tags": [ "descriptor", "attention", "self-optimizing", "CSP", "CALYPSO" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "dataset-dependent (validated on Mg-Ca-H and Be-P-N-O)", "equivariance": "invariant", "architecture": "descriptor", "usesAttention": true, "longRange": false }, { "id": "emff_2025", "type": "node", "category": "Descriptor", "label": "EMFF-2025", "year": 2025, "author": "Wen, Han, Li, Chang, Chu, Chen (Beijing Institute of Technology)", "x": 4580, "y": 900, "desc": "General-purpose deep-potential neural network force field for energetic materials composed of C, H, N, and O. Built on the DeePMD-kit framework with a transfer-learning workflow that fine-tunes a pretrained NNP backbone on a small, targeted set of DFT energies and forces, achieving DFT-level accuracy across 20 different high-energy material (HEM) systems for structural, mechanical, and decomposition properties. Distributed as the EMFF-2025_V1.0.pb model for LAMMPS + DeepMD integration; suitable for MD simulations of HEM systems with 1–5000 atoms, delivering ~30× speedup over baseline LAMMPS via GPU parallel execution.", "githubUrl": "https://github.com/MingjieWen/General-NNP-model-for-C-H-N-O-Energetic-Materials", "paperUrl": "https://doi.org/10.1038/s41524-025-01809-w", "isNew": true, "coverage": [ "energetic materials", "high-energy materials", "explosives", "organic crystals" ], "useCases": [ "explosives MD", "shock simulation", "decomposition chemistry", "mechanical properties" ], "properties": [ "energy", "forces" ], "frameworks": [ "LAMMPS" ], "maintenance": "active", "lastReviewed": "2026-05-22", "trainingData": [ "custom CHNO DFT (HEM systems)" ], "tags": [ "descriptor", "Deep Potential", "DeePMD", "energetic materials", "transfer learning" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "C, H, N, O", "equivariance": "invariant", "architecture": "descriptor", "usesAttention": false, "longRange": false, "numElements": 4 }, { "id": "nep_cg", "type": "node", "category": "Descriptor", "label": "NEP-CG", "year": 2026, "author": "Fan, W. Zhang, Z. Zhang, Xu, Shao, Dong (Bohai University)", "x": 4860, "y": 750, "desc": "Coarse-grained extension of the neuroevolution potential (NEP) framework that generates low-noise training data from the potential of mean force by constraining coarse-grained beads during atomistic simulations and accumulating time-averaged forces. Implemented within GPUMD, NEP-CG reaches training accuracy comparable to atomistic models trained on DFT data while running at hundreds to thousands of ns/day on a single consumer GPU. Demonstrated on liquid water (reproducing densities from 1 bar to 1 GPa with a virial correction for the correct equation of state) and an anisotropic C60 monolayer (capturing directional thermal conductivity). The companion NEP-AACG multiscale variant integrates all-atom and coarse-grained degrees of freedom in a single model, demonstrated for gold nanowire fracture at experimentally relevant strain rates.", "paperUrl": "https://arxiv.org/abs/2603.01234", "isNew": true, "coverage": [ "coarse-grained liquids", "water", "C60 monolayers", "metal nanowires" ], "useCases": [ "coarse-grained MD", "multiscale MD", "mesoscale simulations", "thermal transport" ], "properties": [ "energy", "forces" ], "frameworks": [ "LAMMPS" ], "maintenance": "active", "lastReviewed": "2026-05-22", "trainingData": [ "potential of mean force from atomistic NEP simulations" ], "tags": [ "descriptor", "NEP family", "coarse-grained", "multiscale", "GPUMD" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "—", "equivariance": "invariant", "architecture": "descriptor", "usesAttention": false, "longRange": false, "numElements": null }, { "id": "reaxnet", "type": "node", "category": "Equivariant", "label": "ReaxNet", "year": 2025, "author": "Gao, Yam, Mao, Chen, Chen, Hu (Univ. of Hong Kong / Hong Kong Quantum AI Lab / MattVerse)", "x": 7380, "y": 320, "desc": "Foundation machine learning potential that integrates explicit polarizable long-range electrostatics with an E(3)-equivariant message-passing graph neural network (built on e3nn-jax). Rather than predicting partial charges, ReaxNet employs a physically motivated polarizable charge-equilibration scheme that directly optimizes the electrostatic interaction energy and reproduces polarization responses under external electric fields — capturing long-range effects that elude standard cutoff-based message passing. Trained across the periodic table up to Pu and distributed as a JAX / JAX-MD implementation, it has been demonstrated on mechanical properties, ionic diffusivity in solid-state electrolytes, ferroelectric phase transitions, and reactive dynamics at electrode-electrolyte interfaces.", "githubUrl": "https://github.com/reaxnet/reaxnet", "paperUrl": "https://doi.org/10.1038/s41467-025-65496-3", "isNew": true, "coverage": [ "general materials", "solid-state electrolytes", "ferroelectrics" ], "useCases": [ "materials foundation MLIP", "long-range electrostatics", "ferroelectric phase transitions", "solid-state battery interfaces", "ionic diffusivity" ], "properties": [ "energy", "forces", "stress" ], "frameworks": [ "ASE", "JAX-MD" ], "license": "GPL-3.0", "maintenance": "active", "lastReviewed": "2026-05-25", "tags": [ "equivariant", "polarisable", "long-range-electrostatics", "charge equilibration", "JAX", "e3nn-jax", "foundation model" ], "supportsCharges": true, "supportsSpins": false, "elementsCovered": "all elements up to Pu (Z=94)", "equivariance": "constrained", "architecture": "gnn", "usesAttention": false, "longRange": true, "numElements": 94 }, { "id": "metal53mlp", "type": "node", "category": "Descriptor", "label": "Metal53-MLP", "year": 2026, "author": "X.Y. Li, J. Li, Y.N. Wang et al.", "x": 7100, "y": 550, "desc": "Domain-specific machine-learning interatomic potential for metallic materials covering 53 metallic elements — broad enough to span elemental metals, intermetallics, and multi-principal-element (high-entropy) alloys within a single model. Reaches DFT-level accuracy (energy MAE ≈ 12 meV/atom, force MAE ≈ 144 meV/Å) while staying efficient enough for large-scale molecular dynamics, and reproduces lattice parameters, elastic constants, and equations of state. Validated across four representative alloy problems: negative thermal expansion in orthorhombic Ti-Nb, the Elinvar effect in a Co25Ni25(TiZrHf)50 intermetallic, grain-boundary segregation and high-temperature deformation in the NbTaMoW refractory multi-principal-element alloy, and precipitation pathways.", "paperUrl": "https://doi.org/10.1038/s41524-026-02072-3", "isNew": true, "coverage": [ "metals", "alloys", "high-entropy alloys", "intermetallics" ], "useCases": [ "large-scale alloy MD", "mechanical properties", "phase transformations", "alloy design" ], "properties": [ "energy", "forces", "stress" ], "maintenance": "active", "lastReviewed": "2026-05-26", "trainingData": [ "custom DFT (metallic systems)" ], "tags": [ "descriptor", "metals", "alloys", "high-entropy alloys", "intermetallics", "domain-specific" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "53 metallic elements", "equivariance": "invariant", "architecture": "descriptor", "usesAttention": false, "longRange": false, "numElements": 53 }, { "id": "md_et", "type": "node", "category": "Transformer", "label": "MD-ET", "year": 2026, "author": "Eissler, Korjakow, Ganscha, Unke, Müller, Gugler (BIFOLD / TU Berlin / Google DeepMind)", "x": 7660, "y": 150, "desc": "Molecular-dynamics Edge Transformer: a deliberately minimal adaptation of an off-the-shelf Edge Transformer architecture to interatomic potentials, deployed with neither built-in equivariance nor energy conservation. Rather than encoding physical inductive biases into the network, MD-ET relies on a simple supervised pre-training scheme over ~30 million molecular structures from the QCML database and then fine-tunes for a small number of steps, reaching state-of-the-art accuracy on several molecular-dynamics benchmarks. The paper (\"How simple can you go?\") argues that, as in other ML domains, general-purpose transformer architectures paired with large-scale pre-training can rival specialised, physics-constrained networks for MD.", "paperUrl": "https://doi.org/10.1063/5.0295035", "isNew": true, "coverage": [ "organic molecules", "small molecules" ], "useCases": [ "molecular dynamics surrogate", "transferable pre-training", "fine-tuning to new systems" ], "properties": [ "energy", "forces" ], "frameworks": [ "PyTorch" ], "maintenance": "experimental", "lastReviewed": "2026-05-27", "trainingData": [ "QCML" ], "tags": [ "transformer", "edge transformer", "non-equivariant", "off-the-shelf", "large-scale pretraining" ], "supportsCharges": false, "supportsSpins": false, "elementsCovered": "dataset-dependent (QCML molecular coverage)", "equivariance": "learnt", "architecture": "gnn", "usesAttention": true, "longRange": false } ], "groups": [ { "id": "zone_eq", "type": "group", "label": "Equivariant & Transformers (Accuracy / Foundations)", "x": 50, "y": 50, "width": 3070, "height": 400 }, { "id": "zone_inv", "type": "group", "label": "Invariant & Descriptors (Speed / Scale)", "x": 50, "y": 480, "width": 1620, "height": 500 } ], "edges": [ { "from": "tfn", "to": "nequip", "label": "E(3)", "description": "TFN introduced E(3)-equivariant spherical-harmonic tensor-product convolutions; NequIP applies that machinery to interatomic potentials and explicitly cites TFN-style equivariant message passing as its basis." }, { "from": "tfn", "to": "se3t", "label": "+Attention", "description": "The SE(3)-Transformer generalises self-attention to SE(3)-equivariant inputs using tensor-field-style equivariant filters, i.e. it is TFN with attention bolted on." }, { "from": "nequip", "to": "allegro", "label": "Locality", "description": "Allegro is from the same group (Kozinsky lab) and removes message passing while keeping NequIP's E(3)-equivariant tensor-product machinery, trading non-locality for strict locality and parallel scalability." }, { "from": "nequip", "to": "mace", "label": "Higher Order", "description": "MACE extends NequIP-style E(3)-equivariant message passing with higher-body-order (4-body) messages constructed via Atomic Cluster Expansion, reducing the depth needed for SOTA accuracy." }, { "from": "nequip", "to": "eqv2", "label": "Attention", "description": "Equiformer V2 is an equivariant transformer that uses higher-degree tensor representations of the same E(3)-equivariant family pioneered by NequIP/TFN, reaching SOTA on OC20/OC22 by adding attention to that representation." }, { "from": "se3t", "to": "eqv2", "label": "Refined", "dashed": true, "description": "Equiformer V2 is a refined, higher-capacity equivariant transformer that descends from the SE(3)-Transformer line of equivariant attention." }, { "from": "mace", "to": "grace", "label": "ACE family", "description": "GRACE (Graph ACE) is a foundation-scale graph implementation of the Atomic Cluster Expansion. MACE shares the ACE many-body basis as its mathematical backbone, so GRACE is a parallel ACE-graph development rather than a strict descendant of MACE; the link captures the shared ACE-on-a-graph design." }, { "from": "orb_v2", "to": "orb", "label": "v2 → v3", "description": "Orb-v3 is the direct successor to Orb-v2 from Orbital Materials, sharing the same non-equivariant backbone with refinements to training data and conservativity." }, { "from": "orb", "to": "orbmol", "label": "Adds OMol25", "description": "OrbMol is the molecular variant of Orb-v3 trained on the OMol25 dataset, with the same Orb backbone plus charge/spin conditioning." }, { "from": "pet", "to": "petmad", "label": "MAD pretraining", "description": "PET-MAD applies the Massive Atomic Diversity training recipe to the Point Edge Transformer architecture; PET supplies the unconstrained-equivariance graph-transformer backbone." }, { "from": "eqv2", "to": "esen", "label": "Smooth PES", "description": "eSEN is an equivariant GNN focused on producing a smooth, energy-conserving PES for stable MD; it sits in the same Meta FAIR equivariant lineage as Equiformer V2 but the abstract emphasises smoothness/expressivity rather than naming Equiformer V2 as a parent." }, { "from": "esen", "to": "uma", "label": "Backbone", "description": "UMA is a Mixture-of-Linear-Experts foundation model built on the eSEN equivariant backbone; the UMA paper explicitly identifies eSEN as the underlying architecture." }, { "from": "gemnet", "to": "jmp", "label": "GemNet-OC backbone", "description": "JMP uses a GemNet-OC backbone shared across all training datasets — GemNet-OC supplies the architecture; JMP supplies the joint pretraining strategy on top." }, { "from": "jmp", "to": "uma", "label": "Multi-task", "dashed": true, "description": "JMP demonstrated joint multi-domain pretraining across OC20/OC22/ANI-1x/Transition-1x and is widely framed as the precursor to UMA's universal multi-dataset foundation model." }, { "from": "bpnn", "to": "gap", "label": "Kernels", "description": "GAP (Bartók et al., 2010) was developed as a kernel-based alternative to the Behler-Parrinello descriptor framework, sharing the local-environment / atomic-energy decomposition but replacing the per-element neural network with Gaussian process regression on SOAP descriptors." }, { "from": "bpnn", "to": "ani_1", "label": "Neural Nets", "dashed": true, "description": "ANI-1 builds directly on the Behler-Parrinello high-dimensional neural network framework: per-element atomic networks acting on Behler-style symmetry function descriptors." }, { "from": "ace", "to": "grace", "label": "Graph ACE", "dashed": true, "description": "GRACE is a graph-network implementation of the Atomic Cluster Expansion — explicit in its name (Graph ACE) and in its description as a foundation-scale ACE." }, { "from": "gap", "to": "deepmd", "label": "Neural Nets", "description": "DeepMD replaces GAP-style kernel regression on hand-crafted descriptors with deep neural networks acting on local-frame descriptors, sharing the descriptor + atomic-energy decomposition idea." }, { "from": "deepmd", "to": "dpa2", "label": "Universal Data", "description": "DPA-2 is the second-generation Deep Potential architecture from the DeepMD team, adding attention and multi-task heads on top of the DeepMD framework." }, { "from": "gap", "to": "snap", "label": "Linear Bispectrum", "description": "SNAP (Thompson et al., 2015) fits a linear regression in the same SOAP/bispectrum descriptor space introduced by GAP, replacing GAP's Gaussian-process kernel with a linear model for LAMMPS throughput." }, { "from": "gap", "to": "sgdml", "label": "Kernel ML", "description": "sGDML is a kernel-ridge-regression force field fit in the gradient domain — a kernel-method sibling to GAP rather than a direct descendant; the link captures the shared kernel-ML lineage." }, { "from": "bpnn", "to": "mtp", "label": "Linear Basis", "dashed": true, "description": "MTP shares the Behler-Parrinello atomic-decomposition framework but replaces the symmetry-function + neural-network combination with a linear regression over moment-tensor descriptors." }, { "from": "mtp", "to": "ace", "label": "Many-body Basis", "dashed": true, "description": "ACE generalises the moment-tensor / linear many-body descriptor philosophy of MTP into a complete, systematically improvable many-body basis; Drautz's ACE paper situates itself in the MTP/SOAP descriptor lineage." }, { "from": "gap", "to": "flare", "label": "GP On-the-fly", "description": "FLARE is a Gaussian-process potential built on 2-/3-body or B2/SOAP descriptors with active learning; it is the on-the-fly active-learning successor to the GAP kernel framework." }, { "from": "mtp", "to": "flare", "label": "AL-native", "dashed": true, "description": "FLARE inherits the active-learning-native philosophy that MTP popularised (D-optimal/MaxVol selection); FLARE replaces the linear regression with a Gaussian process and adds Bayesian uncertainty for AL." }, { "from": "flare", "to": "nequip", "label": "Kozinsky Lab", "dashed": true, "description": "Same group: FLARE was developed in the Kozinsky lab and is the immediate kernel-method predecessor to that group's NequIP/Allegro line." }, { "from": "bpnn", "to": "hip_nn", "label": "HDNNP Hierarchy", "description": "HIP-NN extends the Behler-Parrinello high-dimensional NN potential framework with a hierarchical n-body decomposition and residual structure for interpretability and uncertainty estimation." }, { "from": "schnet", "to": "dimenet", "label": "+Angles", "description": "DimeNet/DimeNet++ extend SchNet-style continuous-filter convolutions with directional message passing using spherical Bessel/harmonic bases that explicitly encode bond angles." }, { "from": "ani_1", "to": "ani_1x", "label": "+Active Learning", "description": "ANI-1x is the active-learning extension of ANI-1: query-by-committee active learning grows the training set towards a more chemically diverse organic-molecule sampling." }, { "from": "ani_1x", "to": "ani_1ccx", "label": "+CCSD(T) Transfer", "description": "ANI-1ccx is ANI-1x transfer-learned to a CCSD(T)/CBS reference, yielding a near-coupled-cluster-accuracy organic potential." }, { "from": "ani_1ccx", "to": "ani", "label": "+S, F, Cl", "description": "ANI-2x extends the ANI-1/-1x/-1ccx series by adding S, F, and Cl on top of H, C, N, O." }, { "from": "ani", "to": "aimnet", "label": "AIM Network", "description": "AIMNet was developed in the Isayev group (same lab as ANI) and is a message-passing extension of the ANI atomic-environment-vector concept, propagating those AEVs through self-consistent neighbour updates." }, { "from": "aimnet", "to": "aimnet_nse", "label": "+NSE Charge/Spin", "description": "AIMNet-NSE is the Neural Spin Equilibration variant of AIMNet that handles arbitrary total charge and spin multiplicity through an SCF-like message-passing loop." }, { "from": "aimnet_nse", "to": "aimnet2", "label": "+Long-range", "description": "AIMNet2 extends the AIMNet/AIMNet-NSE charge-aware message-passing line with explicit physics-based long-range electrostatics over 14 elements and ~20M hybrid-DFT data points." }, { "from": "dimenet", "to": "gemnet", "label": "Spherical", "description": "GemNet from the same group (Klicpera/Gasteiger) generalises DimeNet's directional message passing with spherical-harmonic two-hop interactions and quadruplet messages — explicitly framed as the GemNet successor." }, { "from": "dimenet", "to": "alignn", "label": "Line Graph", "description": "ALIGNN augments the atomic graph with a line graph so that bond angles become first-class nodes in the message-passing scheme — a graph-theoretic alternative to DimeNet's spherical-Bessel encoding of the same angular information." }, { "from": "m3gnet", "to": "chgnet", "label": "+Charge", "description": "CHGNet is a charge-aware graph network that extends the M3GNet framework with magnetic-moment / oxidation-state features for redox-active and battery materials." }, { "from": "nequip", "to": "nequix", "label": "Compact NequIP", "description": "Nequix is a compact E(3)-equivariant foundation potential explicitly framed as a simplified NequIP design with equivariant RMSNorm and the Muon optimizer, trained on a small (~100 A100 GPU-hour) budget." }, { "from": "bpnn", "to": "nep89", "label": "Evolution NN", "description": "NEP89 is a Neuroevolution Potential — a per-element neural network on top of local descriptors trained with separable natural evolution strategies — sitting squarely in the Behler-Parrinello HDNNP descriptor + NN family." }, { "from": "eqv2", "to": "liten", "label": "TQA", "description": "LiTEN-FF is an equivariant attention-based network whose Tensorized Quadrangle Attention captures 3- and 4-body interactions in linear time, sitting in the equivariant-transformer lineage that Equiformer V2 popularised." }, { "from": "mace", "to": "liten", "label": "4-body", "dashed": true, "description": "LiTEN-FF's quadrangle attention captures up to 4-body interactions, paralleling the higher-body-order messages that MACE introduced." }, { "from": "sevennet", "to": "sevennet_omni", "label": "Multi-fidelity", "description": "SevenNet-Omni is a multi-fidelity universal foundation extension of the SevenNet equivariant family, using a SevenNet-MF backbone trained across ~15 datasets." }, { "from": "sevennet_omni", "to": "sevennet_nano", "label": "Distillation", "description": "SevenNet-Nano is a distilled lightweight model with SevenNet-Omni as the teacher, delivering an order-of-magnitude speedup at retained accuracy." }, { "from": "mace", "to": "mace_polar1", "label": "Polarisable MACE", "description": "MACE-POLAR-1 is the polarisable extension of MACE adding non-self-consistent atomic charge/spin densities and Fukui equilibration on top of the MACE backbone." }, { "from": "orbmol", "to": "mace_polar1", "label": "OMol25 dataset", "dashed": true, "description": "MACE-POLAR-1 is trained on the same OMol25 dataset that OrbMol uses; the link captures shared training data, not architectural descent." }, { "from": "dpa2", "to": "dpa3", "label": "LiGS", "description": "DPA-3 is the third-generation Deep Potential model, succeeding DPA-2 with a Line Graph Series (LiGS) message-passing scheme that updates bond/angle/dihedral representations." }, { "from": "nep89", "to": "orion", "label": "Organic CHONSP", "description": "ORION is a universal organic (CHONSP) force field built explicitly within the NEP framework that NEP89 popularised, sharing the neuroevolution training and GPUMD deployment." }, { "from": "eqv2", "to": "eqv3", "label": "Scaling", "description": "Equiformer V3 is the next generation in the same Equiformer family, with improvements in efficiency, expressivity, and generality and SOTA on OC20/OMat24/Matbench Discovery." }, { "from": "m3gnet", "to": "matris", "label": "3-body Attn", "description": "MatRIS is an invariant foundation MLIP with separable O(N) attention for three-body interactions, sitting in the M3GNet invariant-GNN lineage with explicit 3-body interactions." }, { "from": "aimnet2", "to": "omni_p2x", "label": "Excited States", "description": "OMNI-P2x is described as an ensemble of MS-ANI-style invariant potentials extended to molecular excited states. It descends from the ANI/AIMNet invariant-organic lineage rather than directly from AIMNet2; AIMNet2's charge/spin handling is a relevant precedent for excited-state conditioning." }, { "from": "ani", "to": "omni_p2x", "label": "MS-ANI", "dashed": true, "description": "OMNI-P2x's backbone is explicitly an ensemble of MS-ANI invariant potentials, i.e. it descends directly from the ANI architectural line." }, { "from": "grace", "to": "grace_off", "label": "GRACE for organics", "description": "GRACE-OFF is a GRACE-architecture MLIP specialised for organic liquids, trained on SPICE v2.0 — the organic-liquids analogue of MACE-OFF for the GRACE family." }, { "from": "petmad", "to": "omnimol", "label": "PET architecture", "dashed": true, "description": "OmniMol adapts the Omnilearned Point-Edge Transformer (PET) architecture — the same PET backbone family that PET-MAD popularised in MLIPs — to molecular dynamics via cross-domain transfer learning." }, { "from": "uma", "to": "omnimol", "label": "OMol25 dataset", "dashed": true, "description": "OmniMol is trained on the OMol25 dataset that UMA helped popularise as a universal molecular benchmark; the link is dataset-level rather than architectural." }, { "from": "painn", "to": "aceff", "label": "TensorNet2", "description": "AceFF is built on TensorNet2, a refined vector-scalar equivariant TensorNet that descends from the PaiNN scalar/vector equivariant message-passing line (TensorNet itself was introduced as a vector-scalar architecture in the PaiNN tradition)." }, { "from": "aimnet2", "to": "aceff", "label": "Charge-aware FF", "dashed": true, "description": "AceFF includes scalar partial-charge features, neutral charge equilibration, and a long-range Coulomb energy term — the same charge-aware MLIP design pattern AIMNet2 popularised for organic chemistry." }, { "from": "mace", "to": "mace_osaka26", "label": "Adds actinides", "description": "MACE-Osaka26 is a multi-domain universal MACE-architecture potential extending the MACE-Osaka series to 97 elements with the new HE26 actinide dataset." }, { "from": "mace_polar1", "to": "mace_osaka26", "label": "MACE family", "dashed": true, "description": "Co-membership in the MACE family; both are MACE-architecture extensions developed concurrently in 2026, but they target different goals (polarisable molecular vs broad-element-coverage materials)." }, { "from": "eqv2", "to": "mlanet", "label": "Dynamic Attn", "description": "MLANet is an efficient equivariant GNN with a geometry-aware dual-path dynamic attention mechanism, sitting in the equivariant-attention MLIP lineage that Equiformer V2 popularised." }, { "from": "nequip", "to": "equiewald", "label": "+Reciprocal Ewald", "description": "EquiEwald embeds an Ewald-inspired reciprocal-space formulation inside an irreducible SO(3)-equivariant message-passing framework — i.e. NequIP-style equivariant message passing extended into reciprocal space." }, { "from": "mace_polar1", "to": "equiewald", "label": "Long-range MLIP", "dashed": true, "description": "Both target long-range physical effects — MACE-POLAR-1 via polarisable charge/spin densities, EquiEwald via reciprocal-space Ewald summation — making them sibling long-range equivariant MLIPs released in the same period." }, { "from": "eqv2", "to": "allscaip", "label": "All-to-All Attn", "description": "AllScAIP is a scalable, energy-conserving, attention-based MLIP that pairs local self-attention with a global all-to-all node attention layer. It sits squarely in the FAIR/Meta equivariant-transformer lineage that Equiformer V2 helped establish, with a stronger global-attention component for long-range interactions." }, { "from": "uma", "to": "allscaip", "label": "OMol25 leaderboard", "dashed": true, "description": "AllScAIP tops the OMol25 leaderboard at release (the same dataset that UMA popularised) while remaining competitive on OMat24 and OC20 — datasets all introduced/used by UMA. The link is shared evaluation/training-data ecosystem." }, { "from": "mace", "to": "mace_mag", "label": "Adds magnetic moments", "description": "MACE-Magnetic extends the MACE framework to magnetic materials by embedding atomic magnetic moments as explicit degrees of freedom alongside positions, with optional spin-orbit coupling." }, { "from": "mace_polar1", "to": "mace_mag", "label": "MACE family", "dashed": true, "description": "Co-membership in the MACE family — both are 2026 MACE extensions to additional physical degrees of freedom (charge/spin densities for POLAR-1, magnetic moments and SOC for Magnetic)." }, { "from": "allegro", "to": "allegro_moe", "label": "Mixture-of-Experts", "description": "Allegro-MoE is a multifidelity Mixture-of-Experts framework explicitly built on the strictly-local E(3)-equivariant Allegro architecture, partitioning the simulation cell into regions with experts of different capacity." }, { "from": "matris", "to": "matris_moe", "label": "MoE Scale", "description": "MatRIS-MoE is a billion-parameter Mixture-of-Experts extension of MatRIS that inserts sparse expert modules around MatRIS's self-attention layer." }, { "from": "mace", "to": "hi_mlip", "label": "+Hessian Training", "description": "Hi-MLIP applies the HINT (Hessian-INformed Training) protocol to MACE-architecture potentials, adding Hessian pre-training, configuration sampling, curriculum learning, and a stochastic projected Hessian loss." }, { "from": "nequip", "to": "hi_mlip", "label": "NequIP backbone", "dashed": true, "description": "Hi-MLIP's HINT protocol is architecture-agnostic and is also demonstrated on NequIP-style equivariant MLIPs as a curvature-aware training enhancement." }, { "from": "painn", "to": "spookynet", "label": "Vector + Charge Attn", "description": "SpookyNet is a self-attention message-passing potential with explicit total-charge and spin-multiplicity tokens. It descends from the equivariant scalar/vector message-passing tradition that PaiNN popularised, with attention and explicit electronic conditioning added." }, { "from": "aimnet_nse", "to": "spookynet", "label": "Charge/Spin Attn", "dashed": true, "description": "Both AIMNet-NSE and SpookyNet were developed around the same time and inject total charge / spin multiplicity as explicit electronic degrees of freedom into the network. AIMNet-NSE uses an SCF-like message-passing loop; SpookyNet uses self-attention with explicit charge/spin tokens. Sibling charge/spin-aware MLIP lineage." }, { "from": "spookynet", "to": "gems", "label": "Biomolecular FF", "description": "GEMS is a SpookyNet-based biomolecular force-field framework, applying SpookyNet to proteins and condensed-phase biomolecular dynamics with top-down/bottom-up sampling." }, { "from": "spookynet", "to": "uma", "label": "Charge-conditioned FF", "dashed": true, "description": "UMA's charge/spin conditioning design pattern — global tokens for total charge and spin multiplicity — was pioneered by SpookyNet (and AIMNet-NSE). The link captures conceptual lineage of charge/spin-aware MLIPs." }, { "from": "spookynet", "to": "mace_polar1", "label": "Charge/Spin", "dashed": true, "description": "MACE-POLAR-1 inherits the explicit charge/spin-density / Fukui-equilibration design pattern that SpookyNet (and AIMNet-NSE) pioneered, applied within the MACE equivariant message-passing backbone." }, { "from": "eqv2", "to": "hienet", "label": "Hybrid Inv-Eq", "description": "HIENet is positioned in the same equivariant-transformer / foundation-model lineage as Equiformer V2 but interleaves cheap invariant message-passing layers with the equivariant ones, achieving SOTA on Matbench Discovery while running substantially faster than EquiformerV2." }, { "from": "sevennet", "to": "hienet", "label": "Faster than SevenNet", "dashed": true, "description": "HIENet benchmarks against SevenNet-l3i5 as a SevenNet-family equivariant baseline and reports ~90% speedup at improved accuracy; the link captures the shared MPTrj-trained equivariant-foundation niche." }, { "from": "mace", "to": "mace_mh1", "label": "Multi-head Replay", "description": "MACE-MH-1 enhances the MACE architecture (stronger element weight sharing, non-linear product-basis tensor decomposition) and adds a multi-head replay post-training scheme that unifies inorganic crystals, surfaces, organic chemistry, and molecular crystals in a single MACE foundation model." }, { "from": "uma", "to": "mace_mh1", "label": "Cross-domain", "dashed": true, "description": "MACE-MH-1 targets the same unified molecular/surface/inorganic foundation niche that UMA established (and uses a similar multi-head, multi-dataset training recipe), but built on the MACE equivariant backbone." }, { "from": "schnet", "to": "mgnn", "label": "+Moments", "description": "MGNN extends the SchNet-style invariant message-passing framework with Cartesian moment representations of 3D molecular graphs, capturing high-order angular structure without explicit equivariant tensor products." }, { "from": "painn", "to": "mgnn", "label": "Vector → Moment", "dashed": true, "description": "PaiNN's scalar/vector message passing is the closest precedent for MGNN's moment-based representations; MGNN generalises that vectorial decomposition into a hierarchy of Cartesian moments while remaining strictly invariant." }, { "from": "painn", "to": "hydragnn", "label": "PaiNN backbone", "description": "HydraGNN-GFM uses a PaiNN-based equivariant scalar/vector message-passing backbone selected via DeepHyper hyperparameter optimization on Frontier; PaiNN supplies the underlying interaction representation around which the multi-task heads are organised." }, { "from": "uma", "to": "hydragnn", "label": "Multi-task foundation", "dashed": true, "description": "HydraGNN-GFM is a multi-task atomistic foundation model jointly trained on 16 first-principles datasets covering 85+ elements, sitting in the same multi-domain foundation-model niche that UMA established for atomistic simulation." }, { "from": "nep89", "to": "qnep", "label": "+Dynamic Charges", "description": "qNEP extends the NEP framework with environment-dependent partial charges represented per-ion by neural networks of the local descriptor vector and explicit Ewald / particle-particle particle-mesh electrostatics, building on the same neuroevolution-potential descriptor backbone that NEP89 popularised." }, { "from": "aimnet_nse", "to": "qnep", "label": "Charge-aware sibling", "dashed": true, "description": "Sibling charge-aware MLIP design: qNEP injects environment-dependent partial charges into the NEP descriptor backbone with explicit electrostatics, while AIMNet-NSE injects total charge / spin states into a message-passing backbone via its Neural Spin Equilibration loop." }, { "from": "petmad", "to": "petmad15", "label": "MAD → MAD-1.5 / r²SCAN", "description": "PET-MAD-1.5 is the direct successor to PET-MAD from the same EPFL lab-cosmo group: same Point Edge Transformer backbone, retrained at the r²SCAN level on the curated MAD-1.5 dataset that extends elemental coverage from 85 to 102 elements with targeted enrichment of molecules, clusters, surfaces, and low-dimensional structures." }, { "from": "mace_osaka26", "to": "petmad15", "label": "Broad-element uMLIP", "dashed": true, "description": "Sibling broad-element universal potentials: MACE-Osaka26 reaches 97 elements within the MACE equivariant family; PET-MAD-1.5 reaches 102 elements within the unconstrained PET transformer family. The link captures shared coverage-of-the-periodic-table goals rather than architectural descent." }, { "from": "nequip", "to": "densnet", "label": "SE(3)-equivariant", "description": "DenSNet uses an SE(3)-equivariant message-passing network to predict density coefficients of an atom-centred Gaussian basis, sitting in the same E(3)/SE(3)-equivariant lineage as NequIP but with electron density (rather than energy) as the central learned quantity, from which a second equivariant network derives total energies for MD." }, { "from": "mace_polar1", "to": "densnet", "label": "Density / electrostatics", "dashed": true, "description": "Both target electronic / electrostatic observables on top of an MLIP: MACE-POLAR-1 augments MACE with non-self-consistent polarisable charge/spin densities for long-range electrostatics, while DenSNet replaces the energy-first paradigm with a density-first one to expose dipole moments, polarizabilities, and IR spectra directly." }, { "from": "petmad", "to": "pet_oam_xl", "label": "OAM training mix", "description": "PET-OAM-XL keeps the unconstrained Point Edge Transformer backbone introduced by PET / PET-MAD and scales it to extra-large capacity, retraining on the OMat24 + sAlex + MPtrj (OAM) data mixture to top the Matbench Discovery leaderboard." }, { "from": "petmad15", "to": "pet_oam_xl", "label": "PET family", "dashed": true, "description": "Sibling 2026 PET releases from the EPFL lab-cosmo group: PET-MAD-1.5 emphasises broad elemental coverage at r²SCAN level, while PET-OAM-XL emphasises raw stability-prediction accuracy via OMat24 pre-training and is positioned as the Matbench Discovery reference model." }, { "from": "nequip", "to": "alphanet", "label": "Local frames", "description": "AlphaNet swaps the spherical-harmonic tensor-product machinery used by NequIP-style E(3)-equivariant MLIPs for atom-centred local frames with learnable geometric transitions, retaining equivariance while reducing per-step cost." }, { "from": "mace", "to": "alphanet", "label": "Frame vs CG tensors", "dashed": true, "description": "AlphaNet targets the same Matbench Discovery / OC2M foundation niche as MACE-style equivariant potentials, replacing Clebsch-Gordan tensor products with cheaper local-frame operations to improve scalability." }, { "from": "ace", "to": "tace", "label": "Cartesian ACE", "description": "TACE reformulates the Atomic Cluster Expansion entirely in Cartesian space, decomposing atomic environments into a hierarchy of irreducible Cartesian tensors that recover the same symmetry-consistent invariants and equivariants as the spherical-harmonic ACE without Clebsch-Gordan coefficients." }, { "from": "grace", "to": "tace", "label": "Foundation-scale ACE", "dashed": true, "description": "Sibling foundation-scale ACE-family MLIPs: GRACE realises a graph implementation of spherical-harmonic ACE while TACE realises an irreducible-Cartesian-tensor implementation of ACE, both released as universal potentials on the Matbench Discovery OAM track." }, { "from": "mace_polar1", "to": "tace", "label": "Tensorial properties", "dashed": true, "description": "Like MACE-POLAR-1, TACE exposes tensorial electronic observables (charges, magnetic moments, external-field response) on top of an ACE-family backbone, additionally including a Latent Ewald Summation module for explicit long-range electrostatics." }, { "from": "orbmol", "to": "transip", "label": "OMol25 dataset", "dashed": true, "description": "Sibling non-equivariant molecular MLIPs trained on the OMol25 dataset: OrbMol uses the Orb graph-network backbone with learnt rotation-invariance via augmentation, while TransIP uses a generic Transformer backbone that acquires SO(3)-equivariance through a latent embedding-space objective rather than via data augmentation." }, { "from": "nequip", "to": "transip", "label": "Latent vs constrained eq.", "dashed": true, "description": "TransIP contrasts with NequIP-style E(3)-equivariant MLIPs by removing all explicit architectural equivariance constraints (no spherical harmonics, no Clebsch-Gordan tensor products) and instead training a generic Transformer to satisfy SO(3)-equivariance through an embedding-space alignment objective." }, { "from": "uma", "to": "transip", "label": "fairchem foundation", "dashed": true, "description": "TransIP builds on top of the fairchem framework introduced alongside Meta FAIR's UMA family, and follows UMA's OMol25-centred molecular foundation-model recipe while replacing the constrained-equivariant Mixture-of-Linear-Experts backbone with an unconstrained Transformer." }, { "from": "mattersim", "to": "mattersim_mt", "label": "+Multi-task heads", "description": "MatterSim-MT extends the MatterSim foundation MLIP from Microsoft Research with multi-task prediction heads for Bader charges, magnetic moments, Born effective charges, and dielectric matrices on top of the same shared 89-element, 0-5000 K, 0-1000 GPa atomistic backbone — keeping MatterSim as the underlying potential-energy-surface predictor." }, { "from": "uma", "to": "mattersim_mt", "label": "Multi-domain foundation", "dashed": true, "description": "Sibling multi-task atomistic foundation models: UMA uses a Mixture-of-Linear-Experts backbone trained jointly across OC20 / OMat24 / OMol25 / ODAC23 / OMC25 with charge/spin conditioning, while MatterSim-MT adds explicit multi-task heads for electronic and tensorial materials properties on top of the MatterSim backbone. Both target unified foundation-model coverage beyond a single PES." }, { "from": "allegro", "to": "fennix_bio1", "label": "Allegro embedding", "description": "FeNNix-Bio1 uses an Allegro strictly-local E(3)-equivariant embedding (scalar + equivariant channels up to l_max=2, three interaction layers, 5.3 Å cutoff) as the neural backbone of its FENNIX hybrid ML + force-field architecture." }, { "from": "mace_off", "to": "fennix_bio1", "label": "Biomolecular FF", "dashed": true, "description": "FeNNix-Bio1 targets the same SPICE-trained, transferable-organic-MLIP niche as MACE-OFF, but layers explicit force-field physics (ZBL repulsion, fluctuating-charge Coulomb, dispersion) on top of an Allegro embedding rather than relying on a pure short-range MACE backbone." }, { "from": "aimnet2", "to": "fennix_bio1", "label": "Charge-aware FF", "dashed": true, "description": "FeNNix-Bio1 inherits the charge-aware MLIP design pattern AIMNet2 popularised for organic chemistry — fluctuating atom-centred partial charges plus an explicit long-range Coulomb energy term — and combines it with explicit dispersion and ZBL repulsion inside the FENNIX hybrid framework." }, { "from": "dtnn", "to": "schnet", "label": "Continuous filters", "description": "SchNet is the direct successor to DTNN from the same Schütt/Tkatchenko/Müller group, replacing DTNN's tensorised pairwise interaction blocks with continuous-filter convolutions while keeping the per-element embedding and energy-decomposition framework." }, { "from": "cgcnn", "to": "megnet", "label": "Crystal Graph", "description": "MEGNet generalises the CGCNN crystal-graph framework to molecules and crystals jointly with global state attributes, sitting in the direct CGCNN crystal-graph lineage." }, { "from": "cgcnn", "to": "alignn", "label": "+Line Graph", "dashed": true, "description": "ALIGNN extends the CGCNN crystal-graph idea by augmenting the atomic graph with a line graph so bond angles become first-class nodes, while preserving the CGCNN-style multi-edge crystal-graph backbone." }, { "from": "cgcnn", "to": "gnome", "label": "Crystal-graph MLIP", "dashed": true, "description": "GNoME's discovery pipeline relies on graph-network MLIPs trained on crystal graphs in the CGCNN tradition, with active learning closing the loop between DFT and graph-network energy prediction." }, { "from": "megnet", "to": "m3gnet", "label": "+3-body", "description": "M3GNet from the same Materials Virtual Lab group extends MEGNet's joint molecule/crystal graph network with explicit three-body interactions and stress prediction for MLIPs." }, { "from": "physnet", "to": "spookynet", "label": "Charge/Dipole", "description": "SpookyNet inherits PhysNet's modular charge/dipole/electrostatic energy decomposition (both from collaborators in the Unke/Müller circle) and adds self-attention with explicit total-charge and spin-multiplicity tokens." }, { "from": "physnet", "to": "aimnet", "label": "Charge-aware MLIPs", "dashed": true, "description": "PhysNet and AIMNet developed concurrently as charge-aware molecular potentials with learnable atomic charges plus Coulomb correction; sibling architectures rather than direct descendants." }, { "from": "cormorant", "to": "nequip", "label": "Equivariant precursor", "dashed": true, "description": "Cormorant pioneered end-to-end SO(3)-equivariant neural networks using Clebsch-Gordan tensor products on irreducible representations; NequIP applies the same equivariant tensor-product machinery to interatomic potentials." }, { "from": "tfn", "to": "cormorant", "label": "Parallel SO(3)", "dashed": true, "description": "Cormorant and Tensor Field Networks were developed in parallel in 2018-2019 as the two early frameworks for SO(3)-equivariant point-cloud/molecular networks using spherical harmonics and Clebsch-Gordan tensor products." }, { "from": "painn", "to": "egnn", "label": "Scalar/vector", "dashed": true, "description": "EGNN and PaiNN were proposed concurrently (both ICML 2021) as lightweight scalar/vector equivariant message-passing alternatives to higher-order tensor methods; sibling architectures rather than direct descendants." }, { "from": "bpnn", "to": "nep_orig", "label": "HDNNP + NES", "description": "NEP sits in the Behler-Parrinello high-dimensional NN potential family — per-element neural networks on local descriptors — but replaces gradient-based training with a separable natural-evolution strategy for raw GPU throughput in GPUMD." }, { "from": "nep_orig", "to": "nep89", "label": "+89 elements", "description": "NEP89 is a direct universal extension of the original NEP framework, scaling the same NEP descriptor + neuroevolution training recipe to 89-element coverage with a curated DFT dataset." }, { "from": "nep_orig", "to": "qnep", "label": "+Dynamic charges", "dashed": true, "description": "qNEP extends the NEP framework with environment-dependent partial charges and explicit Ewald electrostatics, building directly on the NEP descriptor backbone introduced by Fan et al. in 2021." }, { "from": "nep_orig", "to": "orion", "label": "Organic NEP", "dashed": true, "description": "ORION is an organic-chemistry universal force field built within the NEP framework that the original NEP paper introduced." }, { "from": "pfp_orig", "to": "pfp_v8", "label": "v1 → v8", "description": "PFP v8 is the eighth generation of the Preferred Potential, descending directly from the original PFP introduced in 2022 with successive accuracy / coverage / speed improvements and a switch to r2SCAN reference data." }, { "from": "m3gnet", "to": "pfp_orig", "label": "Universal NNP", "dashed": true, "description": "Sibling first-generation universal NNPs released around 2022: PFP and M3GNet were the two earliest universal MLIPs covering >40 elements, developed in parallel for industrial (PFP) and academic (M3GNet) deployment." }, { "from": "pfp_orig", "to": "mattersim", "label": "Universal NNP family", "dashed": true, "description": "MatterSim from Microsoft sits in the same universal-NNP niche that PFP established as the first commercial universal MLIP; the link captures shared positioning rather than architectural descent." }, { "from": "tfn", "to": "equiformer_v1", "label": "+Attention", "description": "Equiformer V1 combines graph attention with TFN-style E(3)-equivariant tensor representations and depthwise tensor-product MLPs; explicit successor to TFN-style equivariant convolutions with attention added." }, { "from": "equiformer_v1", "to": "eqv2", "label": "v1 → v2", "description": "Equiformer V2 is the direct higher-degree successor to Equiformer V1 from the same MIT Atomic Architects / Meta FAIR collaboration, scaling the equivariant attention to higher-degree spherical representations using eSCN convolutions." }, { "from": "scn", "to": "escn", "label": "SO(3)→SO(2)", "description": "eSCN is a direct refinement of SCN that reduces SO(3) tensor-product convolutions to SO(2) by aligning each pair to a common rotation axis, dramatically lowering the cost of high-degree equivariant convolutions." }, { "from": "escn", "to": "eqv2", "label": "Backbone", "description": "Equiformer V2 explicitly uses eSCN convolutions as its high-degree spherical convolutional backbone, citing eSCN as the convolution operator that makes higher-degree representations efficient." }, { "from": "gemnet", "to": "scn", "label": "OC20 lineage", "dashed": true, "description": "SCN succeeds GemNet-OC as Meta FAIR's flagship OC20 model, replacing GemNet-OC's spherical message passing with multichannel spherical signal representations on edge-aligned local frames." }, { "from": "tfn", "to": "so3krates", "label": "+Attention scales", "description": "SO3krates extends TFN-style SO(3)-equivariant tensor-product machinery with invariant scalar attention plus an equivariant filter, enabling long-range-capable equivariant transformers." }, { "from": "spookynet", "to": "so3krates", "label": "Long-range MLIPs", "dashed": true, "description": "SO3krates and SpookyNet share an emphasis on long-range / electronic-structure-aware MLIPs; SpookyNet uses self-attention with charge/spin tokens, SO3krates uses equivariant attention factorised across length scales. Both contributed to the long-range MLIP design space." }, { "from": "painn", "to": "visnet", "label": "Vector-scalar", "description": "ViSNet generalises PaiNN's scalar/vector equivariant message passing with vector-scalar interactive update rules that capture geometric information without explicit higher-order tensor algebra." }, { "from": "deepmd", "to": "dpa1", "label": "+Attention", "description": "DPA-1 introduces an attention layer to the DeepMD descriptor framework as the first attention-based, pretrained Deep Potential model; the direct attention-augmented successor to DeepMD." }, { "from": "dpa1", "to": "dpa2", "label": "v1 → v2", "description": "DPA-2 is the direct second-generation successor to DPA-1 from the DeepModeling team, retaining DPA-1's pretrained descriptor + attention design while expanding to multi-task heads and broader cross-domain pretraining." }, { "from": "mace", "to": "mace_mp0", "label": "+MPtrj foundation", "description": "MACE-MP-0 is the first MACE foundation model: the MACE architecture trained on the Materials Project trajectory dataset (MPtrj) for broadly transferable accuracy across inorganic crystals, surfaces, defects, and molecular crystals." }, { "from": "mace_mp0", "to": "mace_osaka26", "label": "MACE foundation lineage", "description": "MACE-Osaka26 extends MACE-MP-0's universal-MLIP recipe to 97 elements with the new HE26 actinide dataset, sitting directly in the MACE-MP-0 foundation-model lineage." }, { "from": "mace_mp0", "to": "mace_mh1", "label": "Multi-head replay", "dashed": true, "description": "MACE-MH-1 builds on MACE-MP-0's MPtrj-pretrained MACE backbone with stronger element weight sharing, non-linear product-basis tensor decomposition, and a multi-head replay post-training scheme that unifies inorganic, surface, organic, and molecular-crystal data." }, { "from": "mace_mp0", "to": "mace_mag", "label": "Adds magnetism", "dashed": true, "description": "MACE-Magnetic builds on the MACE-MP foundation-model framework, extending the universal MACE backbone with explicit atomic magnetic moments and optional spin-orbit coupling for magnetic materials." }, { "from": "mace_mp0", "to": "mace_polar1", "label": "Adds polarisation", "dashed": true, "description": "MACE-POLAR-1 extends the MACE foundation-model line with non-self-consistent polarisable charge/spin densities and Fukui equilibration; descends from the MACE-MP family rather than the bare MACE architecture." }, { "from": "mace", "to": "mace_off", "label": "Organic FF", "description": "MACE-OFF is the organic-chemistry sibling of MACE-MP-0: the MACE architecture trained on SPICE quantum-chemistry data for transferable molecular dynamics across H, C, N, O, F, P, S, Cl, Br, I." }, { "from": "mace_mp0", "to": "mace_off", "label": "MACE foundation pair", "dashed": true, "description": "MACE-MP-0 (materials, MPtrj, 89 elements) and MACE-OFF (organic, SPICE, 10 elements) are sibling MACE foundation models released by overlapping author teams as the materials and organic universal MACE potentials." }, { "from": "mace_off", "to": "grace_off", "label": "Organic foundation analog", "dashed": true, "description": "GRACE-OFF is the GRACE-architecture analog of MACE-OFF for organic liquids — a direct conceptual sibling using the same SPICE-style organic universal-MLIP recipe within the GRACE rather than MACE family." }, { "from": "m3gnet", "to": "gnome", "label": "Universal materials GNN", "dashed": true, "description": "GNoME and M3GNet are sibling universal materials GNNs from 2022-2023; both target broad-coverage stable-structure prediction, with GNoME scaling active-learning-coupled NNP-driven discovery to 2.2M new crystal structures (380k stable)." }, { "from": "gnome", "to": "mattersim", "label": "Materials discovery", "dashed": true, "description": "MatterSim follows GNoME in the universal materials-discovery foundation-model niche; both leverage active-learning-coupled NNP-driven exploration of crystal-structure space at scale." }, { "from": "aimnet2", "to": "mpnice", "label": "Charge-aware MP", "description": "MPNICE extends the charge-aware message-passing MLIP design popularised by AIMNet2 — invariant local messaging with an explicit iterative charge-equilibration step and long-range Coulomb electrostatics — to an 89-element pretrained model family from Schrödinger." }, { "from": "bpnn", "to": "mpnice", "label": "HDNNP + Qeq", "dashed": true, "description": "MPNICE sits in the Behler-Parrinello high-dimensional NN potential tradition (atomic energy decomposition + local descriptors) and adds an iterative charge-equilibration (Qeq) head and explicit long-range Coulomb electrostatics on top." }, { "from": "ani", "to": "omni_p1", "label": "Multi-fidelity", "description": "OMNI-P1 is the All-in-One ANI (AIO-ANI) network from the Dral group: a single ANI-family invariant-descriptor potential simultaneously trained on multiple quantum-chemical levels (semi-empirical, DFT, CCSD(T)/CBS) so one model spans QC fidelities for organic chemistry." }, { "from": "omni_p1", "to": "omni_p2x", "label": "OMNI family", "description": "OMNI-P2x is the excited-state successor to OMNI-P1 in the Dral-group OMNI series; both share the MS-ANI / AIO-ANI invariant-descriptor backbone, with P1 covering ground-state multi-fidelity learning and P2x extending the design to molecular excited states." }, { "from": "mace", "to": "mace4ir", "label": "+Dipole head", "description": "MACE4IR pairs two MACE E(3)-equivariant networks — one predicting energies/forces, one predicting atomic dipole moments — to deliver foundation-scale molecular infrared spectra; MACE supplies the underlying equivariant message-passing backbone." }, { "from": "mace_off", "to": "mace4ir", "label": "Molecular MACE foundation", "dashed": true, "description": "MACE-OFF and MACE4IR are sibling molecular MACE foundation models trained on quantum-chemistry data covering ~80 elements; MACE-OFF emphasises transferable energies/forces on SPICE, MACE4IR adds an explicit dipole-moment head trained on the QCML dataset for infrared spectroscopy." }, { "from": "mace", "to": "geodite", "label": "No tensor products", "description": "Geodite removes the Clebsch-Gordan tensor products that drive the cost of MACE/NequIP-style E(3)-equivariant message passing, replacing them with cheaper geometric operations while keeping equivariance, and adds ZBL + smooth attenuation priors for a well-behaved potential energy surface." }, { "from": "nequip", "to": "geodite", "label": "Equivariant precursor", "dashed": true, "description": "Geodite is part of the NequIP/MACE E(3)-equivariant message-passing lineage but explicitly removes Clebsch-Gordan tensor products, sitting as a tensor-product-free cousin of NequIP." }, { "from": "mace_mp0", "to": "geodite", "label": "MPtrj foundation", "dashed": true, "description": "Geodite-MP is trained on the same Materials Project trajectory dataset (MPtrj) that defined the MACE-MP-0 universal-MLIP recipe, sharing the foundation-model niche while using a tensor-product-free equivariant backbone." }, { "from": "eqv2", "to": "e2former", "label": "+Wigner 6j", "description": "E2Former is an efficient equivariant transformer in the Equiformer-V2 line that replaces the expensive SO(3) spherical-tensor-product convolution with a Wigner 6j convolution, shifting compute from edges to nodes and lowering tensor-product complexity from O(|E|) to O(|V|) while preserving E(3) equivariance." }, { "from": "se3t", "to": "e2former", "label": "Equivariant transformer lineage", "dashed": true, "description": "E2Former and SE(3)-Transformer share the equivariant-attention-on-spherical-tensor-features design; E2Former modernises that lineage with Wigner 6j convolutions for near-linear-scaling tensor products." }, { "from": "e2former", "to": "ubio_molfm", "label": "E2Former-V2 backbone", "description": "UBio-MolFM is built on E2Former-V2, a direct successor to E2Former that extends the Wigner 6j-based equivariant transformer with Equivariant Axis-Aligned Sparsification (EAAS) and Long-Short Range modelling, and applies it as a foundation MLIP for biomolecular systems." }, { "from": "ubio_molfm", "to": "mace_off", "label": "Bio/organic foundation", "dashed": true, "description": "UBio-MolFM and MACE-OFF both target organic / biomolecular chemistry at near-DFT accuracy; UBio-MolFM specialises in solvated bio-systems up to ~1,500 atoms via E2Former-V2 and the bio-specific UBio-Mol26 dataset, while MACE-OFF covers SPICE-style drug-like molecules." }, { "from": "nequip", "to": "lorem", "label": "Equivariant charges", "description": "LOREM extends the NequIP-style E(3)-equivariant message-passing framework with equivariant — rather than scalar — charges as the long-range message-passing primitive, enabling orientation-dependent long-range physics (electrostatics, dispersion, electron delocalisation) without per-dataset tuning of cutoffs or message-passing depth." }, { "from": "equiewald", "to": "lorem", "label": "Long-range equivariant MLIP", "dashed": true, "description": "Sibling long-range equivariant MLIPs: EquiEwald embeds an Ewald-inspired reciprocal-space formulation inside SO(3)-equivariant message passing, while LOREM keeps the message-passing in real space but promotes the long-range carrier from scalar to equivariant charges. Both target orientation-dependent long-range interactions on top of the NequIP equivariant lineage." }, { "from": "petmad", "to": "lorem", "label": "lab-cosmo", "dashed": true, "description": "Same EPFL lab-cosmo group as the PET / PET-MAD line; LOREM is the lab-cosmo entry into long-range-aware equivariant message-passing MLIPs, complementary to the unconstrained PET transformer backbone." }, { "from": "nequip", "to": "nemp", "label": "Edge → node MP", "description": "NEMP reformulates the edge-equivariant message passing used by NequIP/MACE/Allegro into node-equivariant message passing: a single tensor product per central node against a virtual summed neighbour node, rather than one per neighbour edge, while preserving the same E(3)-equivariant guarantees as NequIP." }, { "from": "allegro", "to": "nemp", "label": "Efficient equivariant MLIP", "dashed": true, "description": "Sibling equivariant-MLIP efficiency optimisations: Allegro removes message passing entirely for strict locality and parallelism; NEMP keeps message passing but moves the equivariant tensor product from edges to nodes, reducing memory and compute by 1–2 orders of magnitude over edge-equivariant baselines." }, { "from": "mace", "to": "nemp", "label": "Cost reduction", "dashed": true, "description": "NEMP targets the same equivariant-MLIP cost bottleneck that MACE-style edge tensor products produce, replacing per-edge tensor products with per-node tensor products against a summed virtual neighbour node while matching MACE-family accuracy across MD17/MD22 and universal-potential benchmarks." }, { "from": "mtp", "to": "polymlp", "label": "Polynomial invariants", "description": "PolyMLP extends the moment-tensor-style polynomial descriptor philosophy of MTP into a systematic polynomial regression over linearly independent O(3) polynomial invariants of the local environment, sitting squarely in the MTP/SNAP linear-descriptor lineage." }, { "from": "snap", "to": "polymlp", "label": "Linear invariants", "dashed": true, "description": "PolyMLP and SNAP are sibling linear/polynomial descriptor MLIPs designed for LAMMPS-scale industrial throughput: SNAP fits a linear regression in the SOAP/bispectrum descriptor space, PolyMLP fits a polynomial regression in O(3)-invariant polynomial features." }, { "from": "ace", "to": "polymlp", "label": "Polynomial basis", "dashed": true, "description": "ACE and PolyMLP are concurrent systematically improvable polynomial-basis MLIP families: both expand the local energy in O(3)-invariant polynomial features and rely on linear/polynomial regression rather than neural networks, with PolyMLP shipped as an OpenKIM portable model driver covering 837 elemental and alloy parameterizations." }, { "from": "bpnn", "to": "acnn", "label": "+Attention", "description": "ACNN sits in the Behler-Parrinello high-dimensional NN potential tradition — per-atom energy decomposition on local descriptors — augmented with stacked multi-head attention modules that act analogously to graph aggregation, extending the effective interaction range beyond the descriptor cutoff." }, { "from": "dpa1", "to": "acnn", "label": "Attention-based descriptor MLIP", "dashed": true, "description": "Sibling attention-augmented descriptor MLIPs: DPA-1 introduces attention into the Deep Potential descriptor framework, while ACNN couples multi-head attention with a generic local descriptor + elemental embedding + MLP fitting head, tuned for crystal-structure-prediction workflows in the CALYPSO ecosystem." }, { "from": "deepmd", "to": "emff_2025", "label": "DeePMD + transfer", "description": "EMFF-2025 is built directly on the DeePMD-kit Deep Potential framework, fine-tuned via transfer learning from a pretrained NNP backbone on a curated DFT energy/force dataset for CHNO high-energy materials; ships as a LAMMPS-compatible Deep Potential model file." }, { "from": "nep_orig", "to": "nep_cg", "label": "+Coarse-grained", "description": "NEP-CG extends the NEP framework from atomistic to coarse-grained interatomic potentials by training on potential-of-mean-force-derived forces from constrained atomistic NEP simulations, retaining the NEP descriptor + neuroevolution training recipe inside GPUMD." }, { "from": "mpnice", "to": "reaxnet", "label": "Equivariant polarizable Qeq", "description": "ReaxNet carries the explicit charge-equilibration-plus-long-range-Coulomb idea of MPNICE onto an E(3)-equivariant message-passing backbone and upgrades it to a polarizable response: instead of predicting partial charges it directly optimizes the electrostatic interaction energy and models induced polarization under external fields, whereas MPNICE uses an invariant message-passing network with an iterative Qeq approximation." }, { "from": "equiewald", "to": "reaxnet", "label": "Long-range equivariant MLIP", "dashed": true, "description": "Sibling equivariant long-range MLIPs: EquiEwald embeds an Ewald-inspired reciprocal-space formulation inside SO(3)-equivariant message passing, while ReaxNet attaches a polarizable charge-equilibration long-range term to an E(3)-equivariant GNN. Both target physically consistent long-range electrostatics across periodic materials beyond the message-passing cutoff." }, { "from": "mace_polar1", "to": "reaxnet", "label": "Polarisable foundation pair", "dashed": true, "description": "MACE-POLAR-1 and ReaxNet are concurrent polarisable foundation potentials built on equivariant backbones: MACE-POLAR-1 augments MACE with a non-self-consistent polarisable field for molecular chemistry (OMol25), while ReaxNet targets materials modelling across the periodic table up to Pu with a polarizable charge-equilibration scheme for electrolytes, ferroelectrics, and electrode interfaces." }, { "from": "orb", "to": "md_et", "label": "Drop equivariance", "description": "MD-ET pushes the non-equivariant, non-conservative design that Orb-v3 popularised for materials to its minimal extreme for molecular dynamics: an off-the-shelf Edge Transformer with neither built-in roto-equivariance nor energy conservation, compensating for the absence of physical inductive biases with large-scale supervised pre-training rather than architectural constraints." }, { "from": "transip", "to": "md_et", "label": "Non-equivariant transformer", "dashed": true, "description": "Sibling non-equivariant Transformer interatomic potentials: TransIP keeps a generic Transformer backbone but recovers SO(3)-equivariance through a latent embedding-space training objective, whereas MD-ET forgoes equivariance (and energy conservation) entirely and instead relies on supervised pre-training over ~30M QCML structures plus light fine-tuning." }, { "from": "mace4ir", "to": "md_et", "label": "QCML pre-training", "dashed": true, "description": "MD-ET and MACE4IR both build on the QCML quantum-chemistry dataset: MACE4IR trains an equivariant dipole head on QCML for infrared spectra, while MD-ET uses ~30M QCML molecular structures as the supervised pre-training corpus for its off-the-shelf edge-transformer potential. The link captures shared training data rather than architectural descent." } ] }