LATTE: an atomic environment descriptor based on Cartesian tensor contractions

TL;DR

LATTE addresses the need for an expressive yet efficient local atomic environment descriptor for MLIPs. It builds Cartesian tensors from neighbor contributions, contracts them into $n$-body scalar terms, and uses trainable localized radial functions, coupled with atomic MLPs, to adapt accuracy and cost. The approach demonstrates competitive accuracy across tungsten, small-molecule benchmarks (rMD17), and a large, multi-species SPICE dataset, while scaling to multi-species, large-system simulations and maintaining computational efficiency. By releasing code in PANNA and enabling integration with ASE, JAX-MD, and LAMMPS, the work offers a practical, flexible option for real-world interatomic potential development. Overall, LATTE provides a scalable, multi-species descriptor with tunable expressivity that competes with fast baselines and approaches state-of-the-art accuracy on challenging datasets, highlighting its potential as a versatile MLIP component.

Abstract

We propose a new descriptor for local atomic environments, to be used in combination with machine learning models for the construction of interatomic potentials. The Local Atomic Tensors Trainable Expansion (LATTE) allows for the efficient construction of a variable number of many-body terms with learnable parameters, resulting in a descriptor that is efficient, expressive, and can be scaled to suit different accuracy and computational cost requirements. We compare this new descriptor to existing ones on several systems, showing it to be competitive with very fast potentials at one end of the spectrum, and extensible to an accuracy close to the state of the art.

Figures (1)

• Figure 1: Validation mean absolute error in energy per molecule (top) and forces per component (bottom) for different descriptors trained on aspirin configurations, as a function of the time needed to run a MD step of 1000 molecules. Each color represents the addition of a new contraction, with increasing number of terms. The starting descriptors only have 2-body terms, indicated as $n$() in the legend, and successive sets start from the previous descriptor (connected with a dashed line) and add an increasing number of terms of the contraction indicated in the legend. The points indicated by arrows and labeled S and M are those used in Table \ref{['tab:rmd17']}.