Atomic Permutationally Invariant Polynomials for Fitting Molecular Force\n Fields Article Swipe
YOU?
·
· 2020
· Open Access
·
· DOI: https://doi.org/10.48550/arxiv.2010.12200
· OA: W3116563477
We introduce and explore an approach for constructing force fields for small\nmolecules, which combines intuitive low body order empirical force field terms\nwith the concepts of data driven statistical fits of recent machine learned\npotentials. We bring these two key ideas together to bridge the gap between\nestablished empirical force fields that have a high degree of transferability\non the one hand, and the machine learned potentials that are systematically\nimprovable and can converge to very high accuracy, on the other. Our framework\nextends the atomic Permutationally Invariant Polynomials (aPIP) developed for\nelemental materials in [Mach. Learn.: Sci. Technol. 2019 1 015004] to molecular\nsystems. The body order decomposition allows us to keep the dimensionality of\neach term low, while the use of an iterative fitting scheme as well as\nregularisation procedures improve the extrapolation outside the training set.\nWe investigate aPIP force fields with up to generalised 4-body terms, and\nexamine the performance on a set of small organic molecules. We achieve a high\nlevel of accuracy when fitting individual molecules, comparable to those of the\nmany-body machine learned force fields. Fitted to a combined training set of\nshort linear alkanes, the accuracy of the aPIP force field still significantly\nexceeds what can be expected from classical empirical force fields, while\nretaining reasonable transferability to both configurations far from the\ntraining set and to new molecules.\n
