Active Learning for Saddle Point Calculation

The saddle point (SP) calculation is a grand challenge for computationally intensive energy function in computational chemistry area, where the may represent transition state. traditional methods need to evaluate gradients of at very large number locations. To reduce expensive computations true gradients, we propose an active learning framework consisting statistical surrogate model, Gaussian process regression (GPR) function, and single-walker dynamics method, gentle accent (GAD), saddle-type states. SP detected by GAD applied GPR gradient vector Hessian matrix. Our key ingredient efficiency improvements method which sequentially designs most informative locations takes evaluations original model these train GPR. We formulate this task as optimal experimental design problem efficient sample-based sub-optimal criterion construct show that new significantly decreases required or force model.