Machine learning prediction of hydrogen atom transfer reactivity in photoredox-mediated C–H functionalization

Abstract

Photoredox-mediated hydrogen atom transfer (HAT) catalysis has reshaped the synthetic strategy of C–H bond functionalization. The rationalization and prediction of HAT reactivity are crucial for the reaction design of photoredox-mediated C–H functionalization. In this work, we report the development of a machine learning model that can predict the HAT barrier of photoredox-mediated HAT catalysis using the physical organic descriptors of the ground state substrate and radical. Based on 2926 DFT-computed HAT barriers of the designed chemical space, the trained AdaBoost model is able to predict the HAT barrier with a mean absolute error of 0.60 kcal mol⁻¹ in the out-of-sample test set. The applicability of the machine learning model is further validated by comparing the prediction against the DFT-computed reactivities on scaffolds and substituents that are not present in the designed chemical space, as well as experimental kinetics data of HAT reaction with the cumyloxyl radical. This work provides a machine learning approach for reactivity prediction from physical organic descriptors and DFT-computed statistics, offering a useful technique that can be directly applied in the experimental designs of photoredox-mediated HAT catalysis.