orbital-materials/orb-models

ORB forcefield models from Orbital Materials

What it solves

Orb-models provides pretrained neural network potentials for atomic simulations, allowing researchers to predict the energy, forces, and stress of atomic systems (like crystals and molecules) without the extreme computational cost of traditional Density Functional Theory (DFT) calculations.

How it works

The project implements a series of pretrained models (Orb-v1 through Orb-v3 and OrbMol) that act as universal force fields. It uses GPU-accelerated graph construction (via ALCHEMI Toolkit-Ops) to represent atomic structures and can be integrated into common simulation workflows using ASE (Atomic Simulation Environment) calculators, TorchSim for batched optimization, or the NVALCHEMI Toolkit for high-performance dynamics. Recent updates include the CoulombModule for long-range electrostatics and D3 dispersion corrections to improve the accuracy of van der Waals interactions.

Who it’s for

It is designed for computational chemists, materials scientists, and physicists who need to perform geometry optimizations or Molecular Dynamics (MD) simulations on diverse molecular and crystalline systems.

Highlights

  • High-Performance Inference: Supports full-model compilation for significant speedups on GPUs and utilizes optimized graph construction for large batches of systems.
  • Diverse Model Suite: Includes specialized models like OrbMol for molecular systems (supporting total charge and spin multiplicity) and Orb-v3 for general atomistic simulations.
  • Confidence Estimation: Orb-v3 models include a confidence head that predicts the likely error (MAE) of the force predictions per atom.
  • Extensibility: Provides a built-in finetuning script to adapt pretrained models to custom datasets stored in ASE sqlite databases.

Related

  • Project
    torchmd/torchmd-net

    A PyTorch-based library providing state-of-the-art neural network potentials for molecular simulations and the tools to train them.

  • Project
    brucefan1983/GPUMD

    A highly efficient, GPU-accelerated molecular dynamics package that enables the training and using of neuroevolution machine-learned potentials for large-scale atomistic simulations.

  • Project
    microsoft/mattersim

    A deep learning atomistic model for simulating bulk materials across various elements, temperatures, and pressures to predict energy, forces, and stress.

  • Project
    materialyzeai/matgl

    A graph deep learning library for materials science that provides various GNN architectures to predict material properties and simulate potential energy surfaces.

  • Project
    facebookresearch/fairchem

    A centralized repository of data and pretrained models for materials science and quantum chemistry, enabling high-performance simulations of molecules and inorganic crystals.