Common data formats and model interfaces for atomistic machine learning

What is it about?

Machine learning for molecular simulations has produced a large ecosystem of software, written in different languages (Python, C++, Fortran, Julia) and built on different frameworks (PyTorch, JAX, scikit-learn). These tools work well individually but exchanging data or models between them requires ad hoc conversion code. Integrating ML models into simulation engines like LAMMPS or GROMACS presents the same problem: each combination of model and simulator needs its own interface. metatensor addresses the data side. It provides a labeled, block-sparse array format designed for atomistic quantities. The labels carry metadata about what each element represents (atom types, angular momentum channels, spatial components), and the format stores gradients (forces, stress) alongside the values they derive from. The data structure works across Python, C, C++, Rust, and Fortran through a shared C library. metatomic addresses the model side. It wraps a trained ML model together with metadata describing its inputs, outputs, and capabilities into a portable archive. A simulation engine that supports the metatomic interface can load any compliant model without knowing how the model works internally. The paper also describes the ecosystem built around these two libraries: metatrain for training workflows, featomic for computing atomic descriptors, and integrations with LAMMPS, i-PI, ASE, PLUMED, eOn, and other simulation tools.

Featured Image

Photo by Dayne Topkin on Unsplash

Photo by Dayne Topkin on Unsplash

Why is it important?

The atomistic ML field has a fragmentation problem. Each group builds models in their preferred framework, and using someone else's model in your simulation requires writing interface code specific to that model-simulator pair. The number of required interfaces grows as the product of models and simulators, which does not scale. metatensor and metatomic reduce this to a sum: each model implements the metatomic interface once, each simulator implements it once, and all combinations work. This is the same pattern that made file formats like HDF5 and protocols like MPI successful in scientific computing. The practical consequence: a researcher can train a model with metatrain, export it as a metatomic archive, and run it in LAMMPS, i-PI, or eOn without writing any interface code. The model's metadata ensures that the simulator uses it correctly (right units, right neighbor list settings, right output quantities). The libraries are designed for long-term maintainability, with CI testing across platforms, semantic versioning, and backwards-compatible data formats.

Perspectives