Local Spin-Center Graph Neural Networks for Predicting Intersystem Crossing Rates in Radical Pairs

Self- and cross-reactions of organic peroxy radicals, RO2 + R'O2, are important pathways in atmospheric oxidation because they can form low-volatility accretion products contributing to secondary organic aerosol. The reaction is generally believed to proceed through a tetroxide (RO4R') intermediate. The fragmentation of this tetroxide leads to an alkoxy radical pair, which is initially formed in a triplet-coupled state. Access to the singlet surface through intersystem crossing (ISC) controls the branching between recombination versus radical-propagating channels. However, direct quantum-chemical evaluation of ISC rates for large radical-pair ensembles remains computationally demanding. Here, we develop a local spin-center graph neural network (GNN) for predicting kISC in alkoxy radical pairs. The representation is built around the two spin-bearing oxygen atoms and includes local O-C-X spin-center geometry, element-resolved second-shell atoms, p-frame node/edge projections, and minimal inter-radical core contacts. The best model uses a multichannel additive scheme, where the T1 → S1, T1 → S2, T1 → S3, and T1 → S4 rates are predicted separately, and the total rate is reconstructed as log10(∑iki). For a combined data set consisting of both previously computed radical pairs, and a new set of more complex and atmospherically representative species, the model achieved R2 = 0.907, MAE = 0.320, and RMSE = 0.464 for total log10(kISC). Descriptor-importance analysis showed that ISC rates are mainly controlled by local O-C-X spin-center distortion, followed by element-specific second-shell geometry and p-frame orientation. The proposed GNN provides an accurate and interpretable framework for high-throughput prediction of spin-dependent pathways in atmospheric radical chemistry.