MLIP
mlip is a Python library for building, training, and deploying machine learning interatomic potentials in JAX.
It provides tools for:
Multiple architectures including MACE, NequIP, ViSNet and eSEN; built in a modular way that makes building new models easy
Highly customizable dataset preprocessing for training and inference
Train or fine-tune MLIP models, on a single or multiple accelerators in parallel (even scalable across hosts)
Molecular dynamics, energy minimization, and transition state search with multiple backends
Ultra-fast (batched) inference and MD simulations enabled by JAX-MD backend
Global charge conditioning, treatment of long-range interactions, and training on Hessian labels
Getting Started¶
If you’re new to mlip, we recommend starting here:
Set up mlip and dependencies.
Tutorials and practical workflows.
Detailed API documentation.
Upgrade from v1 to v2.
Note
The mlip library is under active development.