Transferable and Transparent Energy Decomposition-based Machine Learning Models for Computing Accurate Reaction Energetics

We present a transferable, interpretable, and modular machine-learning framework that enhances the accuracy of density functional theory (DFT) reaction energies using physically meaningful energy-decomposition descriptors. Reaction energies computed at the DFT level with standard basis sets are first decomposed into chemically intuitive contributions - such as kinetic and potential energy - which are then used to train a library of linear regression (LR) models. This includes a general-purpose model that reduces mean absolute percentage errors (MAPE) relative to gold standard CCSD(T)/CBS reference values by up to 63% compared to uncorrected DFT across extended benchmark sets. In parallel, a series of specialized LR models provide improved accuracy for specific reaction classes. A random forest (RF) classifier dynamically selects the optimal model for each case, pushing accuracy further and achieving MAPE reductions of up to 123%, all while maintaining full interpretability. In a rigorous out-of-distribution stress test on the WCCR10 dataset - containing transition-metal complexes absent from training - both the general LR model and the RF/LR pipeline retain robust performance. Compared to typical neural network models, which may suffer from generalization issues outside the training set, our framework shows stable performance outside its training domain.