A modular software ecosystem for atomistic machine learning

The metatensor ecosystem is a collection of low-level libraries and tools that can be used to develop, train and use machine learning models to simulate materials and molecules at the atomic scale.

You can find practical examples of how to use metatensor tools in your simulation on the metatensor section of the Atomistic Cookbook.

Reference

If you use one of the tools in the metatensor ecosystem for your research, please cite:
"Metatensor and metatomic: foundational libraries for interoperable atomistic machine learning", arXiv:2508.15704

metatensor ecosystem logo

Core projects

The three core pillars of the metatensor ecosystem

Self-describing sparse tensor data format for atomistic machine learning and beyond

Code Docs

Atomistic machine learning models you can use everywhere for everything

Code Docs

Train, fine-tune, and manipulate machine learning models for atomistic systems

Code Docs

Libraries

Software libraries that build on top of the metatensor ecosystem to provide custom building blocks for all your machine learning tasks.
If you are developing a library that uses metatensor or metatomic, please contact us so we can add it to this list!

sphericart

A library for the efficient calculation of real spherical harmonics and their derivatives in Cartesian coordinates, including fast CUDA kernels.

The sphericart.metatensor module outputs the spherical harmonics using metatensor format.

Code Docs

featomic

Computing representations for atomistic machine learning, including multiple variants of SOAP and the long range LODE features; as well as tools to construct custom equivariant features.

All of featomic's outputs are returned in metatensor format.

Code Docs

torch-spex

A library for the spherical expansions of atomic neighbourhoods, using PyTorch to provide GPU acceleration and automatic differentiation.

The spex.metatensor module outputs the spherical expansion using metatensor format.

Code Docs

torch-pme

Particle-mesh based calculations of long-range interactions in PyTorch

The torchpme.metatensor module contains code that follows the metatomic API.

Code Docs

vesin

Neighbor list calculator for molecular systems.

The vesin.metatomic module contains tools to compulte neighbor lists for metatomic models.

Code Docs

Simulation tools

Simulation tools that can use metatomic models to run molecular dynamics, Monte Carlo, enhanced sampling, and more!

LAMMPS-metatomic

A fork of the LAMMPS molecular dynamics simulator, providing integration with metatomic.

You can use pair_style metatomic to run simulations using metatomic models to predict energies and forces.

Code Docs

i-PI

A universal force engine for advanced molecular simulations, including path integral molecular dynamics, replica exchange, and more.

You can run simulations using metatomic models though the metatomic driver.

Code Docs

PLUMED

Enhanced sampling and free energy methods for molecular simulations.

You can use the METATOMIC action to use collective variables based on metatomic models.

Code Docs

chemiscope

An interactive structure-property explorer for atomistic simulation data.

chemiscope.explore() can use metatomic models to generate features to describe structures and atomic environments.

Code Docs

eOn

Software for long time-scale dynamics of atomistic systems.

eOn can use metatomic models to run simulation with machine learning potentials.

Code Docs

Torch-Sim

TorchSim is an open-source atomistic simulation engine for the MLIP era, built on PyTorch and designed for batchablity.

Torch-Sim can use metatomic models to run simulations.

Code Docs

ASE

The Atomistic Simulation Environment is a set of tools and Python modules for setting up, manipulating, running, visualizing and analyzing atomistic simulations.

You can use MetatomicCalculator to run simulations using metatomic models to predict energies and forces.

Code Docs

Machine learning models

Ready to use machine learning models for diverse tasks in atomistic simulations.

PET-MAD

A lightweight universal interatomic potential for advanced materials modeling.

Code Article

ShiftML

A deep learning model for chemical shieldings in molecular organic solids including anisotropy.

Code Article

FlashMD

Long-stride, universal prediction of molecular dynamics

Code Article