Machine Learning Force Fields (MLFFs) enable modeling of chemical systems combining <i>ab initio</i> accuracy with the efficiency of mechanistic force fields. Despite the great success of ML methods, extending their applicability to larger molecules poses a challenge partly due to the rapid growth of the dimensionality of the descriptor. State-of-the-art descriptors include non-essential degrees of freedom or neglect long-range interactions by imposing a cut-off radius. Thus, finding a way to accurately describe both short- and long-range interactions without significantly increasing the descriptor size would drastically advance ML modeling.<br/><br/>In the present work, we propose a global descriptor optimization scheme that can be efficiently used to model large and flexible molecules. To achieve this, we developed a robust descriptor reduction methodology by employing sensitivity analysis. We apply the developed methodology to various systems of interest, including one supramolecular complex and units of four major types of biomolecules. Our results show that global descriptors can be reduced to less than 30% of their original size without losing accuracy for molecules containing 50 – 150 atoms. Moreover, employing the optimized descriptor results in a 2-4x increase in the efficiency of the modeling. Analysis of the relevant features demonstrates that some features included in local descriptors are not required for accurate models, while certain long-range features (even beyond 10 Å) can be crucial. The developed technique provides a robust descriptor optimization method opening an avenue for efficient and accurate modeling of large molecules.

Optimizing descriptors for accurate machine learning force fields of large molecules

Fall 2022 abstracts, posters, and presentations will launch in mid-September.

ACS Fall 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.