featomic-torch v0.6.0

This is the first fully public release of featomic-torch! This release brings to everyone the following features:

• C++ and Python TorchScript bindings to featomic, making all calculators accessible from TorchScript models.

• Integration of the TorchScript calculators with metatensor-torch-atomistic, using the System class from this package as a system provider, and integrating with neighbor lists provided by the simulation engine through metatensor.

• Automatic integration with (Py)Torch automatic differentiation system. If any of the inputs requires gradients, then featomic-torch will compute them, store them, and integrate them with a custom backward function on the calculator output.

• Re-export of Python tools for Clebsch-Gordan tensor products from featomic, in a TorchScript-compatible way.

featomic v0.6.0

This is the first fully public release of featomic! This release brings to everyone the following features:

• Multiple atomistic features calculators with a native implementation:

  • SOAP spherical expansion, radial spectrum, power spectrum and spherical expansion for pairs of atoms;
  • LODE spherical expansion;
  • Neighbor list;
  • Sorted distances vector;
  • Atomic composition.

• All the calculator outputs are stored in metatensor's TensorMap objects. This allow to both store the features in a very sparse format, saving memory; and to store different irreducible representations (for SO(3) equivariant atomsitic features)

• Most of the calculators can compute gradients with respect to positions, cell or stress, storing them in the gradient() of metatensor's TensorBlock.

• All the native calculators are exposed through a C API, and accessible from multiple languages: Rust, C++ and Python.

• Interface to mutliple system providers, and a way to define custom system providers in user code. The following system providers are supported from Python: ASE (https://wiki.fysik.dtu.dk/ase/); chemfiles (https://chemfiles.org/); and PySCF (https://pyscf.org/)

• Python-only calculators, based on Clebsch-Gordan tensor products to combine equivariant featurizations. This includes

  • PowerSpectrum, able to combine two different spherical expansions
  • EquivariantPowerSpectrum, the same but producing features both invariant and covariant with respect to rotations
  • DensityCorrelations to compute arbitrary body-order density correlations;
  • ClebschGordanProduct, the core building block that does a single Clebsch-Gordan tensor product.

• Python tools to define custom atomic density and radial basis functions, and then compute splines for the radial integral apearing in SOAP and LODE spherical expansions. This enables using these native calculators with user-defined atomic densities and basis functions.

featomic-torch-v0.6.0-rc2

Release featomic-torch v0.6.0-rc2

featomic-torch-v0.6.0-rc1

Release featomic-torch v0.6.0-rc1

featomic-v0.6.0-rc2

Release featomic v0.6.0-rc2