Predicting intersystem crossing rate constants of alkoxy-radical pairs with structure-based descriptors and machine learning

Abstract

Peroxy radicals (RO₂) are ubiquitous intermediates in many oxidation processes, especially in the atmospheric gas phase. The recombination reaction of two peroxy radicals (RO₂ + R′O₂) has been demonstrated to lead, via several steps, to a triplet complex of two alkoxy radicals: ³(RO˙⋯R′O˙). The different product channels of RO₂ + R′O₂ reactions thus correspond to different reactions of this triplet complex. Of particular interest to atmospheric chemistry is the intersystem crossing (ISC) to the singlet state, which enables the recombination of the two radicals to an ROOR′ peroxide with considerably lower volatility than the original precursors. These peroxides are believed to be key contributors to the formation of secondary organic aerosol (SOA) particles, which in turn contribute to both air pollution and radiative forcing uncertainties. Developing reliable computational models for, e.g., RO₂ + R′O₂ branching ratios requires accurate estimates of the ISC rate constants, which can currently be obtained only from computationally expensive quantum chemistry calculations. By contrast, machine learning (ML) methods offer a faster alternative for estimating ISC rate constants. In the present work, we create a dataset with 98 082 conformations of radical pairs and their corresponding rate constants. We apply three ML models—random forest (RF), CatBoost (CB), and a neural network (NN)—to predict ISC rate constants from triplet to singlet states. Specifically, the models predict k_ISC(T₁ → S_i) for i = 1–4 and the cumulative k_ISC(T₁ → S_n), in alkoxy radical pairs, using only molecular geometry descriptors as inputs. All ML models achieved a mean absolute error (MAE) on our test set within one order of magnitude and a coefficient of determination R² > 0.82 for all rate constants. Overall, the ML prediction matches the quantum chemical calculations within 1–2 orders of magnitude, providing a fast and scalable alternative for quantum chemical methods for ISC rate estimation.