Parametrization of Reactive Potential using Genetic Algorithm and Machine Learning Techniques

Multi-scale modeling of composites spans many length scales ranging from atomistic to the continuum level. At the lower length scales, classical molecular dynamics (MD) provides insight into the failure mechanisms in the fiber, matrix, and the interphase region that forms during processing. MD requires selection of an accurate potential energy model to describe the system. Proper usage of the potential energy model requires parameter optimization of the model’s numerous parameters with respect to experimental geometric and energetic constraints. Parametrization of a potential model can be problematic due to its time-consuming and labor-intensive nature. The goal of this work is to develop a machine learning inspired evolutionary parametrization technique to decrease the time cost and diminish the human intervention required in the parametrization process. The evolutionary genetic algorithm is employed to optimize the parameters of the ReaxFF interatomic potential. Several augmentations are implemented using machine learning techniques. The utilization of an artificial neural network as a surrogate for the ReaxFF potential is considered. Changes to the genetic algorithm are incrementally benchmarked for accuracy and time cost with respect to a simple zinc-oxide model. Utilizing the 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.