Compare · 22 models
Molecular simulation and materials models you can run online
Computational engines and machine learning potentials for atomistic simulation, molecular dynamics, crystal property prediction, and generative materials discovery across chemistry and materials science.
What molecular simulation and materials models do
Atomistic simulations compute the energetic, structural, and dynamical properties of molecular and solid-state systems. Conventional electronic structure approaches like density functional theory provide high physical fidelity but impose computational costs that limit simulation sizes and timescales. Classical molecular dynamics engines address larger systems by evaluating empirical force fields over time-stepped trajectories.
Machine learning interatomic potentials and atomistic foundation models bridge this computational gap. By learning from quantum mechanical reference data, they predict total potential energies, atomic forces, partial charges, and stress tensors orders of magnitude faster than ab initio calculations. Related architectures predict self-consistent exchange-correlation functionals, evaluate electron densities, or generate novel crystalline lattices conditioned on target physical properties.
Compared
- AIMNet2 (B97-3c, 2025)isayevlabA neural network interatomic potential calculating molecular energies, forces, and charges at the B97-3c level.Check termsOn demand
- AIMNet2 ωB97M-D3isayevlabNeural network interatomic potential trained at the ωB97M-D3 level of theory for fast molecular simulations.Check termsOn demand
- AIMNet2-NSEisayevlabA neural network interatomic potential for open-shell chemistry, radicals and non-singlet molecular states.Check termsOn demand
- AIMNet2-PdisayevlabA neural network interatomic potential predicting energies, forces, and charges for palladium-containing systems.Check termsOn demand
- AIMNet2-rxnisayevlabA neural-network interatomic potential for predicting energies, forces, and pathways in organic chemical reactions.Check termsOn demand
- MACE-MP-0mace-foundationsAn equivariant message-passing potential trained on the Materials Project for inorganic crystal simulation.Check termsOn demand
- MACE-MPA-0mace-foundationsMACE foundation model predicting energy, forces, and charges for inorganic materials and molecules.Check termsOn demand
- OpenMMChodera Lab / StanfordGPU-native molecular dynamics simulation engine with a scriptable Python API for biomolecular systems.Commercial OKOn demand
- AmberToolsAmber consortiumA molecular simulation suite providing parameterisation, solvation, and free-energy evaluation for molecular systems.Check termsNot yet
- BOA — Basis Overlap Architecturesciai-labAn equivariant graph neural network predicting molecular ground-state electron densities using basis overlap matrices.See its pageNot yet
- Equiformer v3mirror-physicsAn SE(3)-equivariant graph transformer predicting energies and forces for molecules and crystalline materials.See its pageNot yet
- GROMACSGROMACS consortiumPerforms molecular dynamics simulations of large biomolecular systems and membrane proteins.Commercial OKNot yet
- MatterGenmicrosoftA diffusion model that generates novel, stable inorganic crystal structures with user-defined target properties.See its pageNot yet
- MatterSimmicrosoftAn atomistic foundation model for predicting material properties across varied temperatures and pressures.See its pageNot yet
- OC25facebookCatalysis model simulating explicit-solvent and catalyst systems for electrochemistry and energy applications.See its pageNot yet
- OMat24facebookInorganic material property prediction models using EquiformerV2 and eSEN architectures.See its pageNot yet
- OMC25facebookPredicts properties of organic molecular crystals for chemistry and materials science researchers.See its pageNot yet
- OMol25facebookPredicts molecular properties and energies for chemistry and materials science researchers.See its pageNot yet
- OrbMolOrbital-MaterialsA molecular foundation-model interatomic potential predicting energies, forces, and properties across chemical systems.See its pageNot yet
- Skala 1.0microsoftA neural exchange-correlation functional for DFT delivering hybrid accuracy at semi-local computational cost.See its pageNot yet
- Skala 1.1microsoftA deep-learning exchange-correlation functional for DFT combining semi-local speed with hybrid accuracy.See its pageNot yet
- UMAfacebookA family of universal atomic graph neural networks for chemistry and materials science simulations.See its pageNot yet
How to choose
For classical molecular dynamics of large biomolecular and condensed-phase systems, GROMACS and OpenMM evaluate time-dependent coordinate trajectories from PDB inputs. GROMACS is configured for large macromolecular complexes and membrane proteins embedded in explicit solvent. OpenMM provides a GPU-native molecular dynamics engine driven via a scriptable Python API. ambertools functions as a parameterisation, solvation, and analysis toolchain for the Amber force field family, including MM/PBSA and MM/GBSA free-energy calculations.
For molecular systems requiring quantum mechanical accuracy, interatomic potentials depend on chemical composition and electronic state. The AIMNet2 family operates across 14 main-group elements (H, B, C, N, O, F, Si, P, S, Cl, As, Se, Br, I). AIMNet2 (B97-3c, 2025) and AIMNet2 ωB97M-D3 predict potential energy surfaces, forces, and Hessians for closed-shell molecules. For radicals, triplets, or excited states, AIMNet2-NSE accepts explicit charge and spin multiplicity inputs. AIMNet2-Pd swaps arsenic parameterisation for palladium to simulate organometallic catalysts, while AIMNet2-rxn restricts elemental coverage strictly to neutral closed-shell H, C, N, and O species for batched nudged elastic band reaction pathway searches. Alternatively, OrbMol predicts conservative and direct forces across organic and metal-organic systems with explicit charge and spin inputs using a 6 Å cutoff radius, and OMol25 provides eSEN and AllScAIP checkpoints via fairchem for molecular property prediction.
For solid-state crystal lattices and inorganic materials, MACE-MP-0 and MACE-MPA-0 apply equivariant message passing to Materials Project and combined Materials Project/Alexandria datasets respectively, covering most of the periodic table. OMat24 implements EquiformerV2 and eSEN architectures to evaluate energies, forces, and stress tensors for cell relaxation. Equiformer v3 uses graph attention transformers with SwiGLU-S^2 activations evaluated across datasets including OMat24 and MPtrj. OMC25 focuses specifically on organic molecular crystals such as pharmaceutical polymorphs, while UMA provides a mixture-of-experts model trained across five separate benchmarks (OC20, ODAC23, OMat24, OMC25, OMol25). OC25 specifically models catalyst-adsorbate interfaces in explicit solvent for electrochemistry. For broad thermodynamic conditions, MatterSim functions across varied temperatures and pressures, whereas MatterGen uses diffusion to generate novel inorganic crystal unit cells of up to 20 atoms.
For electronic structure calculations, Skala 1.0 and Skala 1.1 replace standard exchange-correlation terms within iterative self-consistent field DFT frameworks (such as PySCF), delivering hybrid-level accuracy at semi-local meta-GGA computational scaling. When the objective is predicting ground-state electron density directly rather than potential energy surfaces, BOA uses basis function overlap matrices within an equivariant graph neural network.
Questions
Can I run these models without a GPU or installing anything?
8 of the 22 models here run on our servers from their page: open one, fill in the inputs and press Run. Those marked "Not yet" are listed with their licence and documentation but have no loader here yet.
Which models support open-shell systems and radicals?
AIMNet2-NSE and OrbMol are specifically designed to model open-shell species, transition states, and radicals. Both accept explicit total charge and spin multiplicity as user inputs.
How do MACE-MP-0 and MACE-MPA-0 differ?
Both are equivariant message-passing potentials developed by mace-foundations that require at least 16 GB of GPU memory. MACE-MP-0 was trained on inorganic crystal structures from the Materials Project, whereas MACE-MPA-0 was trained on a combined dataset of the Materials Project and Alexandria datasets to broaden compositional coverage.
How do Skala 1.0 and Skala 1.1 differ from interatomic potentials?
Interatomic potentials directly map 3D nuclear coordinates to energies and atomic forces without computing electronic states. Skala 1.0 and Skala 1.1 are neural network exchange-correlation functionals designed to run inside iterative self-consistent field (SCF) density functional theory calculations on an integration grid.
When should MatterGen be used instead of MatterSim or OMat24?
MatterGen is a generative diffusion model that invents new inorganic crystal structures of up to 20 atoms per unit cell conditioned on target material properties. MatterSim and OMat24 are evaluative models that predict energies, forces, or mechanical properties for existing, input atomic structures.
Written from each model's own documentation, as summarised on its page here. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.