Prediction of Activation Energies of Organic Molecules With at Most Seven Non‐Hydrogen Atoms Using Quantum‐Chemically Assisted <scp>ML</scp>

In this study, a hybrid machine learning (ML) approach is presented for accurately predicting activation energies (E_a) of gas-phase elementary reactions involving organic compounds with up to seven non-hydrogen atoms. Given the importance of activation energies in reaction studies and modeling, ML composite models were created that effectively integrate molecular descriptors with semi-empirical and single energy density functional theory (DFT) calculations. The dataset, containing 300 randomly selected elementary gas-phase reactions, was assembled using accurate DFT (ωB97X-D3/def2-TZVP) values for activation energies E_a from a database alongside semi-empirical computations. For accurate predictions, this approach required the inclusion of both physical organic and geometric/empirical descriptors in the training procedure. The best two ML models demonstrated efficient E_a prediction capability, achieving a mean absolute error (MAE) of 1.314 kcal mol^-1 and R² of 0.992 (Model 3) and (MAE) of 1.949 kcal mol^-1 and R² of 0.979 (Model 2) in validation tests. Notably, this performance approaches the threshold of "chemical accuracy" of 1 kcal mol^-1. Model's 3 robustness was tested across the reaction types present in the dataset, demonstrating its ability in properly predicting activation energies, which is critical for the study and optimization of chemical processes.