Unlocking XAT/HAT selectivity: a machine learning-guided discovery of governing physicochemical principles

Abstract

Halogen-atom transfer (XAT) has emerged as a powerful strategy for activating carbon–halogen (C–X) bonds to efficiently generate carbon-centered radicals. However, realizing the full synthetic potential of XAT requires mechanistic clarity, as the factors governing selectivity between XAT and the frequently competing hydrogen-atom transfer (HAT) pathway remain poorly understood. To address this challenge, we introduce a data-driven computational framework that couples high-accuracy density functional theory (DFT) calculations with machine learning (ML) to accurately predict HAT/XAT selectivity across a diverse panel of carbon radical initiators. Specifically, the random forest algorithm demonstrated the highest predictive accuracy for the relative activation barriers (ΔΔG^‡). Mechanistic analysis, leveraging feature importance ranking and multivariate linear regression, revealed two decisive physicochemical descriptors for reaction control: the bond dissociation energy (BDE) of the organic halide substrate and the ionization potential (IP) of the initiating radical species. The predictive power and generalizability of this principle are confirmed through validation studies involving synthetically relevant boryl, nitrogen-centered, and thiol radicals. This work not only showcases the power of ML-guided mechanistic discovery in physical organic chemistry but also provides a robust, predictive platform for the rational design and optimization of highly selective halogen-atom transfer processes.