Metatensor’s goals

At its core, metatensor provides tools to efficiently store and manipulate sparse arrays and their associated metadata. You can learn more about this in the core classes overview. With the creation of metatensor, we have three main use cases in mind:

1. provide an exchange format for the atomistic machine learning ecosystem, making different players in this ecosystem more interoperable with one another and enhancing collaboration: see Exchanging data;

2. make it easier and faster to develop new machine learning representations, models and algorithms: see Defining custom models;

3. run large scale simulations using machine learning interatomic potentials, with fully customizable potentials, directly defined by the researchers running the simulations: see Running atomistic simulations;

Exchanging data

First, metatensor is a format to exchange data between different libraries in the atomistic machine learning ecosystem. There is currently an explosion of libraries and tools for atomistic machine learning, implementing new representation, new models, and advanced research methods. Unfortunately each one of these libraries lives mostly separated from the others, resulting in a lot of duplicated effort. With metatensor, we want to provide a way for these libraries to communicate with one another, by giving everyone a lingua franca, a way to share data and metadata.

Illustration of machine learning workflows, going from some input data to a prediction. Metatensor enables the creation of workflows mixing different libraries together in new ways.

This goal is enabled by multiple features of metatensor: first, metatensor allows storing data coming from many different sources, without requiring to first convert the data to a specific format. Currently, we support data stored inside numpy arrays, torch tensor (including tensors on GPU or other accelerators), as well as arbitrary user-defined C, C++, and Rust array types. A second part of this goal is achieved by also storing metadata together with the data, communicating between libraries exactly what is stored in the different arrays. We also store both data and gradients of this data with respect to arbitrary parameters together, enabling for example training of models using energy, forces and virial. Finally, we also make sure that the data storage is as efficient as possible and can exploit the inherent sparsity of atomistic data, in particular in gradients.

As a developer a library in the atomistic machine learning ecosystem, you can provide conversion functions to and from metatensor metatensor.TensorMap (either inside your own code or in a small conversion package) to enable using your library in conjunction with the rest of the metatensor ecosystem!

libraries using metatensor

The following libraries use metatensor either as input, output or both:

• featomic: a library computing physics-inspired representations of atomic systems, the computed representations are given in metatensor format;

• torch_spex: pure PyTorch implementation of spherical expansion representation, with GPU and learnable representations support, which outputs to metatensor format;

• metatrain: an end-to-end training and evaluation library for models based on metatensor;

• Q-stack: library of pre- and post-processing tasks for Quantum Machine Learning; can output some of its data in metatensor format;

Defining custom models

The second objective of metatensor is to provide functionalities to be a tool for developing new models. While it is possible to use metatensor to only exchange data between libraries (and immediately convert everything to library-specific formats); we also provide tools to operate directly on metatensor data. This enable models to handle sparse data and have low memory consumption; as well as keeping rich metadata around for easier debugging and understanding of the model behavior.

One part of these tools is the set of low-level operations we provide as part of the Python interface to metatensor. By using combining multiple operations, you can build custom machine learning models, using data and representations coming from arbitrary metatensor-compatible libraires in the ecosystem. Using these operations allow you to keep your data in metatensor format across the whole ML pipeline; ensuring the metadata is kept up to date with the data, and gradients are automatically updated to stay consistent with the values.

Another part of these tools is the learning utilities, which provide high level building blocks for machine learning models, with API similar to PyTorch or scikit-learn. These blocks enable you do define and train models with a few lines of code and a familiar API.

Warning

The learning utilities are still an early work in progress, with a lot more building blocks to be included.

Where similar functionalities is provided by different packages

Package	Core data class	Operations	Machine learning models facilities
numpy	numpy.ndarray	numpy.pow()	scikit-learn
torch	torch.Tensor	torch.pow()	torch.nn.Module, torch.utils.data.Dataset
metatensor	metatensor.TensorMap	metatensor.pow()	metatensor.learn.nn.ModuleMap, metatensor.learn.Dataset

Running atomistic simulations

One particularly interesting class of machine learning model for atomistic modelling is machine learning interatomic potentials (MLIPs). Using the capacities provided by the first two goals of metatensor, researchers should be able to created and train such MLIPs and customize various parts of the model.

The final objective of metatensor is to allow using these custom models inside large scale molecular simulation engines. To do this, we integrate metatensor with TorchScript, and use the facilities of TorchScript to export the model from Python and then load and execute it inside the simulation engine. Have a look at the Atomistic applications section for more information!

Different steps in the workflow of running simulations with metatensor. Defining a model, training a model and running simulations with it can be done by different users; and the same metatensor-based model can be used with multiple simulation engines.