Overview
TorchSim is a next-generation open-source atomistic simulation engine for the MLIP (machine-learning interatomic potential) era. By rewriting the core primitives of atomistic simulation in PyTorch, it achieves significant acceleration of popular machine-learning potentials. It supports MACE, Fairchem, SevenNet, ORB and MatterSim MLIP models, plus classical Lennard-Jones, Morse and soft-sphere potentials, and provides NVE/NVT/NPT molecular dynamics, gradient-descent and FIRE relaxation, Swap Monte Carlo, and integration with ASE, Pymatgen and Phonopy.
TorchSim is a simulation engine that consumes machine-learning potentials (an AI-related tool) rather than a machine-learning model itself. It is released under MIT; the paper is AI for Science 1, 025003 (2025).
Key Features
- Torch-native, batchable atomistic simulation engine with GPU acceleration and automatic batching.
- Supports MACE, Fairchem, SevenNet, ORB, MatterSim and classical potentials.
- Provides NVE/NVT/NPT molecular dynamics, FIRE relaxation and Swap Monte Carlo.
Use Cases
- Run batched molecular dynamics and relaxation with machine-learning potentials on GPUs.
- Act as a simulation engine connecting multiple MLIPs with ASE/Pymatgen/Phonopy workflows.
How to Use
Install with pip install torch-sim-atomistic and use with MLIP models such as MACE. See https://torchsim.github.io/torch-sim/.
Sources and verification
This entry describes capabilities documented by the linked project sources. Catalog edits are not evidence of a successful run, compatibility test or scientific performance evaluation.
Read the editorial and verification policy