Simulating phase transformation of materials at the atomistic scale requires the knowledge of saddle points on the potential energy surface (PES). In the existing first-principles saddle point search methods, the requirement of a large number of expensive evaluations of potential energy, e.g. using density functional theory (DFT), limits the application of such algorithms to large systems. Thus, it is meaningful to minimize the number of functional evaluations as DFT simulations during the search process. Furthermore, model-form uncertainty and numerical errors are inherent in DFT and search algorithms. Robustness of the search results should be considered. In this paper, a new search algorithm based on Kriging is presented to search local minima and saddle points on a PES efficiently and robustly. Different from existing searching methods, the algorithm keeps a memory of searching history by constructing surrogate models and uses the search results on the surrogate models to provide the guidance of future search on the PES. The surrogate model is also updated with more DFT simulation results. The algorithm is demonstrated by the examples of Rastrigin and Schwefel functions with a multitude of minima and saddle points.
Data-driven surrogate models for LTI systems via saddle-point dynamics
