Atomistic applications#

While metatensor is a generic sparse data container able to store data and metadata for multiple scientific fields, it comes from the field of atomistic machine learning and as such offer some additional facilities for defining and using machine learning models applied to atomistic systems.

Metatensor provides tools to build your own models (in the form of operations), define new models architectures and export models you just train to use them in arbitrary simulation engines. If you want to train existing architectures with new data or re-use existing trained models, look into the (work in progress!) metatensor-models project instead.

💡 Tutorials

Learn how to define your own models using metatensor, and how to use these models to run simulation in various simulation engines.

📋 Standard models outputs

Understand the metadata convention for specific models outputs, such as the potential energy.

Python Python API reference

Read the documentation for all the classes and functions related to atomistic models in Python.

Documentation for version 0.2.1

C++ C++ API reference

Read the documentation for all the classes and functions related to atomistic models in C++.

Documentation for version 0.2.1

Overview#

All the model facilities in metatensor are based on PyTorch and in particular TorchScript; and as such are part of the metatensor-torch package. This allow users to define new models with Python code (as a custom torch.nn.Module instance), train the models from Python, and export them to TorchScript. The exported model can then be loaded into a C++ simulation engine such as LAMMPS, GROMACS, etc. and executed without relying on a Python installation.

Metatensor provides code for using atomistic systems as input of a machine learning model with metatensor.torch.atomistic.System, and exporting trained models with metatensor.torch.atomistic.MetatensorAtomisticModel. Such models can make predictions for various properties of the atomistic system, and return them as a dictionary of metatensor.torch.TensorMap, one such tensor map for each property (i.e. energy, atomic charges, dipole, electronic density, chemical shielding, etc.)