|
Abstract:
|
ReaxFF is a widely used simulator based on reactive force field method to simulate material properties. In order to obtain the correct physics from the ReaxFF simulations, a crucial step is to optimize the force field parameters by minimizing the discrepancy between simulation results and those from more accurate but time-consuming approaches such as Quantum mechanics-based approaches. To optimize the force field parameters in ReaxFF, there are two challenges. First, there are infeasible regions in the parameter space which are unknown. So, it is crucial to first identify the infeasible regions and then perform optimization only within the feasible regions. Identifying infeasible regions in computer simulations is an important classification problem for many scientific applications, but the development of binary classification received scant attention in computer experiment literature. Second, when analyzing the computer experiments, the commonly used Gaussian process (GP) models are computationally intractable for massive data with high-dimensionality, which is common for complex system such as ReaxFF. To tackle both problems, we introduce a two-stage procedure which incorporates
|
