Precise estimation of activation energies in gas-phase chemical reactions via artificial neural network

Various machine learning (ML) models are presented in this study, aiming to forecast the barrier heights (BHs) of gas-phase chemical reactions. The input features utilized in six distinct models were obtained from the structural and thermodynamic attributes of molecules, encompassing enthalpy, topological indices, and Morgan fingerprints derived from SMILES, using a dataset consisting of 5040 decomposition reaction records sourced from the Gas Phase Organic Chemistry database. Evaluating the effectiveness of the models included the application of essential metrics such as coefficient of determination, mean absolute error, and root mean square error. It is worth noting that artificial neural networks outperform the other models in this regard. Then we utilized Morgan fingerprints of different dimensions as inputs for the neural network models and conducted training with varying numbers of hidden layers. This endeavor led to slight improvements in the performance of gas-phase decomposition reactions, resulting in an average determination coefficient of 0.965 and a mean absolute error of 0.079 eV. Subsequently, the model was subjected to retraining using a comprehensive dataset comprising a wide range of chemical reactions. The results indicate that the artificial neural network approach has the capacity to generalize and adjust to a wider range of chemical reactions.