Automated ReaxFF parametrization using machine learning

Molecular dynamics (MD) simulation requires an accurate potential energy function to describe atomic interactions of interest. Optimization of the function's numerous parameters is often time-consuming and labor-intensive. In this study, a machine learning inspired evolutionary parametrization technique using the genetic algorithm is developed to decrease the time required to optimize the parameters of the ReaxFF interatomic potential. An artificial neural network is used as a surrogate for the ReaxFF potential to reduce computational time. Changes to the genetic algorithm are incrementally benchmarked for accuracy and time cost with respect to a moderately complex zinc-oxide model to find superior operators for ReaxFF parametrization. It is found that utilizing an artificial neural network significantly boosted performance, as measured by the final total error and the rate of decrease of total error with respect to time. The double-Pareto probability density based crossover operator and a multiple standard deviation based Gaussian mutation scheme outperform their counterparts. The computational time cost to achieve the same level of accuracy relative to manual training is decreased from months to days.