What is it about?
Simulations of atoms require an algorithm that computes what are the energy and the forces for each set of positions of the atoms. These can be computes using very precise physical models, but those are computationally expensive. Alternatively, one can train a machine learning model on data and then use it in place of the slow physical model. In this paper, we show that an informed and accurate encoding of the information about the relative positions of atoms can be used together with very simple regression tools and yield accuracy which is on-par with the one yielded by very complex, and computationally more expensive, regression models such as deep neural networks.
Photo by Pietro Jeng on Unsplash
Why is it important?
This article showcases how the ``atomic cluster expansion'' descriptor is able to compactly encode information about the relative positions of atoms, and how this can enable accurate force and energy predictions even while using very simple and fast regression models. We also show that in principle the descriptor can be made even more compact through PCA feature selection, for a variety of systems.
Read the Original
This page is a summary of: Compact atomic descriptors enable accurate predictions via linear models, The Journal of Chemical Physics, June 2021, American Institute of Physics,
You can read the full text:
Github Repository of Python Package
Ridge-regression Atomistic Force Fields in PYthon Use this package to train and validate force fields for single- and multi-element materials. The force fields are trained using ridge regression on local atomic environment descriptors computed in the "Atomic Cluster Expansion" framework.
The following have contributed to this page