Machine learning analysis of dynamic‐dependent bond formation in trajectories with consecutive transition states

Abstract

AbstractDynamic motion often controls selectivity in reactions featuring two consecutive potential‐energy transition states. Here we report density functional theory (DFT)‐based direct dynamics trajectories and machine learning classification analysis for cyclopentadienone dimerization and a N2 extrusion reaction leading to semibullvalene. These reactions have consecutive transition states, and there is dynamic selectivity that determines which of two possible C‐C bonds is formed after the first transition state. For cyclopentadienone dimerization with a bispericyclic first transition state, machine learning analysis using transition‐state based features provided >90% trajectory classification accuracy, but only using AdaBoost and random forest algorithms. Many other relatively sophisticated machine learning algorithms showed poor accuracy despite the obvious motion responsible for selectivity. Feature importance analysis confirmed that the sigmatropic rearrangement vibrational motion in the bispericyclic transition state provides prediction of which of the second C‐C bonds is dynamically formed. For the reaction leading to semibullvalene, machine learning analysis provides solid accuracy for classifying trajectories and predicting which C‐C bond is formed and which C‐C bond is broken immediately after N2 ejection. Like the cyclopentadienone dimerization reaction, machine learning feature importance analysis showed that the sigmatropic rearrangement vibrational motion in the N2 extrusion transition state determines which C‐C bond is formed and which is broken. Surprisingly, machine learning struggles to predict which trajectories undergo a subsequent [3,3] sigmatropic rearrangement process, which isomerizes equivalent forms of semibullvalene.