Pattern recognition in chemical dynamics

Abstract

We have used pattern recognition methods to predict the outcome of classical trajectory calculations directly from the initial conditions, bypassing the need to integrate the trajectories numerically. Two schemes have been used, a nearest neighbor method (NN) and an adaptive digital learning network (DLN). In each, prediction success greater than 80% is achieved. Both schemes use a suitable training set of known trajectories to teach recognition of initial conditions and prediction of trajectory results. For the present example (two atoms recombining on a solid surface), NN achieves a saving in computer time of a factor of roughly 500 over integrating the trajectories. For DLN this factor is about 100.