Learn more about OrbMol

OrbMol is a suite of quantum-accurate machine learning models for molecular predictions. Built on the Orb-v3 architecture, OrbMol provides fast and accurate calculations of energies, forces, and molecular properties at the level of advanced quantum chemistry methods.

The models combine the transferability of universal potentials with quantum-level accuracy, making them suitable for a wide range of applications in chemistry, materials science, and drug discovery.

OMol and OMol-Direct

  • Training dataset: OMol25 (>100M calculations on small molecules, biomolecules, metal complexes, and electrolytes)
  • Level of theory: ωB97M-V/def2-TZVPD with non-local dispersion; solvation treated explicitly
  • Inputs: total charge & spin multiplicity
  • Applications: biology, organic chemistry, protein folding, small-molecule drugs, organic liquids, homogeneous catalysis
  • Caveats: trained only on aperiodic systems → periodic/inorganic cases may not work well
  • Difference: OMol enforces energy–force consistency; OMol-Direct relaxes this for efficiency

OMat

  • Training dataset: OMat24 (>100M inorganic calculations, from Materials Project, Alexandria, and far-from-equilibrium samples)
  • Level of theory: PBE/PBE+U with Materials Project settings; VASP 54 pseudopotentials; no dispersion
  • Inputs: No support for spin and charge. Spin polarization included but magnetic state cannot be selected
  • Applications: inorganic discovery, photovoltaics, alloys, superconductors, electronic/optical materials
  • Caveats: magnetic effects may be incompletely captured

OrbMol supports the following molecular structure formats:

  • .xyz - XYZ coordinate files
  • .pdb - Protein Data Bank format
  • .cif - Crystallographic Information File
  • .traj - ASE trajectory format
  • .mol - MDL Molfile
  • .sdf - Structure Data File

All formats are automatically converted internally for processing.

Single Point Energy: Upload a molecular structure and select a model to calculate energies and forces.

Molecular Dynamics: Run time-dependent simulations to observe molecular behavior at different temperatures and conditions.

Relaxation/Optimization: Find the minimum-energy configuration of your molecular structure.

Each tab provides specific parameters you can adjust to customize your calculations.

All models are based on the Orb-v3 architecture, the latest generation of Orb universal interatomic potentials.

Key features:

  • Graph neural network architecture
  • Equivariant message passing
  • Multi-task learning across different quantum chemistry methods
  • Billions of training examples across diverse chemical spaces
  • Sub-kcal/mol accuracy on test sets

Citation: If you use OrbMol in your research, please cite the Orb-v3 paper and the relevant dataset papers (OMol25/OMat24).

OrbMol — Quantum-Accurate Molecular Predictions

Welcome to the OrbMol demo! This interactive platform allows you to explore the capabilities of our quantum-accurate machine learning models for molecular simulations.

Quick Start

Use the tabs above to access different functionalities:

  1. Single Point Energy: Calculate energies and forces for a given molecular structure
  2. Molecular Dynamics: Run MD simulations using OrbMol-trained potentials
  3. Relaxation / Optimization: Optimize molecular structures to their minimum-energy configurations

Simply upload a molecular structure file in any supported format (.xyz, .pdb, .cif, .traj, .mol, .sdf) and select the appropriate model for your system.

Model Selection Guide

Choose OMol/OMol-Direct for:

  • Organic molecules and biomolecules
  • Drug-like compounds
  • Metal-organic complexes
  • Molecules in solution
  • Systems where you need to specify charge and spin

Choose OMat for:

  • Inorganic crystals and materials
  • Periodic systems
  • Bulk materials and alloys
  • Solid-state compounds

Explore the accordions on the left to learn more about each model's capabilities, training data, and limitations.

Try an Example

To get started quickly, navigate to any of the calculation tabs above and try one of these examples:

  • Single Point Energy: Upload a small molecule to see energy and force predictions
  • Molecular Dynamics: Run a short simulation at 300K to observe thermal motion
  • Relaxation: Optimize a distorted structure to find its equilibrium geometry