Software development for atomistic learning

We dedicate considerable effort to develop software to make advanced atomistic simulations easier to use, more efficient and more accurate. In particular, we are currently working on the following projects:

  • chemiscope: a graphical tool for the interactive exploration of materials and molecular databases, as well as a library of re-usable components useful to create new interfaces.
  • scikit-matter: a collection of scikit-learn compatible utilities that implement methods born out of the materials science and chemistry communities.
  • librascal: a versatile and scalable fingerprint and machine learning code
  • metatensor: a library to manipulate arrays that are associated with rich metadata information, such as those that commonly appear in representations for atomistic machine learning. Think numpy ndarray or pytorch Tensor equipped with extra metadata for atomic — and other particles — systems.
  • rascaline (this is a fork of a repo developed in our group): a library to compute representations for atomistic machine learning
  • nice: a set of tools designed for the calculation of invariant and covariant atomic structure representations.
  • i-pi (this is a fork of the main i-PI repo containing some experimental features developed in our group): a universal force engine interface written in Python, designed to be used together with an ab-initio, empirical or machine-learning force field (or a combination of these!) to run molecular dynamics simulations, including nuclear quantum effects.
  • sphericart, a multi-language library for the efficient calculation of the spherical harmonics and their derivatives in Cartesian coordinates.

You can follow our open-source work directly on the COSMO github page.