Dynamic Effects and Machine-Learning Transition State Theory

Transition state theory is chemistry’s most important quantitative method for the calculation of rates and qualitative framework for the understanding of rates. Some flaws and limitations of transition state theory were apparent at its beginning, while others have become apparent in recent years from a growing number of reactions found to exhibit “dynamic effects,” that is, experimental kinetic observations that cannot be predicted or understood from statistical rate theories. Trajectory methods can often account for dynamic effects but they intrinsically provide very little insight, and each new prediction requires a new set of trajectories. This seminar will describe a new form of transition state theory that uses machine learning to divide transition states in phase space into regions that lead to specific products or transition state recrossing. In this way, machine learning is used to define transmission coefficients for each class of product or recrossing. In simplest form, this process requires an initial set of trajectories, but this set can be six to ten times smaller than a normal set, with equivalent precision, and further predictions can be made without additional trajectories. In an advanced form, the interactive use of machine learning and trajectories allows accurate quantitative predictions with only a small portion of the normally required trajectories. The seminar will describe the application of this process to a series of complex organic reactions where experimental data is available and where conventional and variational transition state theories fail. The results make detailed predictions of temperature effects on product ratios and rates, and provide insight into the origin of recrossing and trajectory selectivity in reactions. On a larger scale, the results define the onset of chaos as the predictability of trajectory outcomes declines with time.