Targeted TPS Shooting Using Computer Vision to Generate Ensemble of Trajectories

This study presents a transition path sampling (TPS) procedure to create an ensemble of trajectories describing a chemical transformation from a reactant to a product state, augmented with a computer vision technique. A 3D convolutional neural network (CNN) sorts the slices of the TPS trajectories into reactant or product state categories, which aids in automatically accepting or rejecting a newly generated trajectory. Furthermore, information about the geometrical configuration of each slice enables one to calculate the percentage of reactant and product states within a specific shooting range. These statistics are used to determine the most appropriate shooting range and, if needed, to improve a shooting acceptance rate. To test the automated 3D CNN TPS technique, we applied it to collect an ensemble of the transition paths for the rate-limiting step of the Morita-Bayliss-Hillman (MBH) reaction.