Prediction of Standard Combustion Enthalpy of Organic Compounds Combining Machine Learning and Chemical Graph Theory: A Strategy

The prediction of thermochemical properties such as the standard enthalpy of combustion is essential for the design and evaluation of energetic materials. In this study, the prediction of this thermochemical property is proposed through a QSPR strategy that combines machine learning and chemical graph theory. The data set consisted of 3477 organic compounds. SMILES codes were used for each molecule to construct their molecular graphs, from which topological indices such as Estrada, Wiener, and Gutman, as well as centrality measures, were calculated. These descriptors served as predictors in supervised learning models, with tree-based ensemble models showing the best performance. The best-performing model, random forest, achieved the following metrics on the test set: R2 = 0.9810, MAE = 287.5988 kJ·mol–1, MAPE = 0.1048, RMSE = 551.9050 kJ·mol–1, and RMSLE = 0.1933. Interpretability analysis using SHAP confirmed that the Estrada and Gutman indices were the most influential variables in the predictions. In addition, the same random forest model was trained using 210 molecular descriptors obtained from RDKit, yielding slightly better metrics: R2 = 0.9927, MAE = 142.2272 kJ·mol–1, MAPE = 0.0484, and RMSE = 342.0464 kJ·mol–1, and RMSLE = 0.1172. Moreover, specific models were developed for different families of compounds, achieving R2 ≈ 0.99 in all cases. Finally, a clustering analysis using the K-Means algorithm in the space defined by the topological indices enabled the identification of latent molecular patterns, providing a novel framework for organizing and analyzing chemical space. This work demonstrates the potential of combining supervised and unsupervised learning methods with chemical graph theory to enable accurate, robust, and scalable prediction of thermochemical properties such as combustion enthalpy.