Gradient-Based Optimization of ReaxFF Parameters Using Pytorch for the Study of Silica Precipitation

Abstract

Silica precipitation is a subject of big interest since it occurs in a wide variety of environmental and industrial processes. Even though there are many advances in atomistic simulation research of different forms of silica, the mechanism of silica precipitation has not been fully understood. We propose to study the following process using reactive force-field method (ReaxFF). Despite being a classical force field, ReaxFF can achieve quantum chemical accuracy once the optimal potential coefficients are found. However, the fitting of ReaxFF parameters is a challenge due to the complex functional form of the potential. Several techniques have been proposed to solve this problem, such as evolutionary algorithms, Monte Carlo methods, and simulated annealing. The stochastic nature of these methods requires millions of error evaluations to fit the parameters, which results in excessive optimization times. Recent advances in machine learning made it possible to drastically speed up the process by utilizing the gradient of the potential. In this work, the gradient-based optimization of reactive force-field parameters using Pytorch was performed. We have implemented ReaxFF potential as a Pytorch model. The model’s performance was validated against existing ReaxFF implementations. ReaxFF parameters were fitted to the dataset, which comprised 15345 geometries calculated using a long-range corrected hybrid functional 𝜔B97XD3.