ML Atomistic
Foundational libraries enabling ML models and simulation engines to communicate
The problem
Atomistic machine learning has a fragmentation problem. Every group builds models in their preferred framework (PyTorch, JAX, etc.), and integrating them with simulation engines (LAMMPS, GROMACS, ASE) requires custom glue code for each combination. A model trained in one framework cannot be used in another without rewriting the integration layer.
The approach
Metatensor provides a shared data format (TensorMap) for representing atomistic ML data: atomic descriptors, model predictions, training labels. The format is framework-agnostic and supports sparse, block-structured tensors with metadata about which atoms and properties each block describes (Bigi et al. 2025).
Metatomic wraps trained models into a standard interface that simulation engines can call. One integration per engine, instead of one per model-engine pair.
My contributions
As part of EPFL labCOSMO, I contribute to the metatensor ecosystem with a focus on interoperability with existing simulation codes. This includes integrations with eOn (saddle point searches) and testing infrastructure for the metatomic model wrapping layer.
Code
- metatensor – Contributor at EPFL labCOSMO
- metatomic – Contributor
- vesin – Neighbor list library