INDEEDopt: a deep learning-based ReaxFF parameterization framework

Mert Y Sengul,Yawei Gao,Tirthankar Dasgupta,Nadire Nayir,Yao Song,Adri C T Van Duin,Ying Hung

doi:10.1038/s41524-021-00534-4

Abstract

Empirical interatomic potentials require optimization of force field parameters to tune interatomic interactions to mimic ones obtained by quantum chemistry-based methods. The optimization of the parameters is complex and requires the development of new techniques. Here, we propose an INitial-DEsign Enhanced Deep learning-based OPTimization (INDEEDopt) framework to accelerate and improve the quality of the ReaxFF parameterization. The procedure starts with a Latin Hypercube Design (LHD) algorithm that is used to explore the parameter landscape extensively. The LHD passes the information about explored regions to a deep learning model, which finds the minimum discrepancy regions and eliminates unfeasible regions, and constructs a more comprehensive understanding of physically meaningful parameter space. We demonstrate the procedure here for the parameterization of a nickel–chromium binary force field and a tungsten–sulfide–carbon–oxygen–hydrogen quinary force field. We show that INDEEDopt produces improved accuracies in shorter development time compared to the conventional optimization method.

Highlights

Atomistic-scale insights have been critical to understanding the dynamical evolution of chemically reactive systems and have created a demand for the development of computational chemistry techniques
The high-dimensional parameter space is sampled by means of an initial design algorithm called Orthogonal-maximin Latin Hypercube Design (OMLHD)
The OMLHD algorithm can generate parameter combinations within ranges specific to each parameter that are multidimensionally uniformly distributed by reducing the pairwise correlation and maximizing the distance between parameters[33]

Summary

INTRODUCTION

Atomistic-scale insights have been critical to understanding the dynamical evolution of chemically reactive systems and have created a demand for the development of computational chemistry techniques. Because the intra- and interatomic interactions are tuned by these parameters, they are optimized to reproduce reference values with reasonable accuracy before moving to production simulations These reference values form a force field training (FFtraining) set, which is composed of, but not limited to, molecular properties (e.g. bond lengths, bond angles, charges, and energies), and/or chemical reactions from as simple as bond breaking/ formation to more complex such as vacancy dynamics, ion diffusion of reference systems. A typical ReaxFF training set is composed of molecular and condensed phase parts, which are separated into groups such as the equation of states, the heat of formations, etc It is very likely for a force field parameter set to get stuck in a local minimum that precisely reproduce reference properties corresponding to a specific part of the force field while poorly fitting to the remaining.

RESULTS

METHODS

CODE AVAILABILITY