What is featomic

Featomic is a library for the efficient computing of representations for atomistic machine learning also called “descriptors” or “fingerprints”. These representation can be used for atomistic machine learning (ML) models including ML potentials, visualization or similarity analysis.

There exist several libraries able to compute such structural representations, such as DScribe, QUIP, and many more. Featomic tries to distinguish itself by focussing on speed and memory efficiency of the calculations, with the explicit goal of running molecular simulations with ML potentials. In particular, memory efficiency is achieved by using the metatensor to store the structural representation. Additionally, featomic is not limited to a single representation but supports several:

representation	description	gradients
Spherical expansion	Atoms are represented by the expansion of their neighbor’s density on radial basis and spherical harmonics. This is the core of representations in SOAP (Smooth Overlap of Atomic Positions)	positions, strain, cell
SOAP radial spectrum	Atoms are represented by 2-body correlations of their neighbors’ density	positions, strain, cell
SOAP power spectrum	Atoms are represented by 3-body correlations of their neighbors’ density	positions, strain, cell
LODE Spherical Expansion	Core of representations in LODE (Long distance equivariant)	positions
Sorted distances	Each atom is represented by a vector of distance to its neighbors within the spherical cutoff	no
Neighbor List	Each pair is represented by the vector between the atoms. This is intended to be used as a starting point for more complex representations	positions
AtomicComposition	Obtaining the stoichiometric information of a system	positions, strain, cell