A Molecular Descriptor-Based Analysis of Organic Compounds and Designing New Promising Candidates by Machine Learning

A data-driven approach using Python has been employed to mine new energetic materials with higher heats of sublimation (ΔH). A data set of 307 energetic compounds with experimental ΔH values was used to train machine learning models, including linear, random forest, gradient boosting, and extra trees, which yielded high R-squared values of 0.95–0.98. The descriptors ATS 1s and mZagreb1 are found to have a high Pearson correlation with the ΔH, while SHapley Additive exPlanations (SHAP) impact analysis reveals that ATS 1s and Wpath have a significant impact on model performance. Using these optimized models, 705 new organic compounds are designed with predicted ΔH as high as 170 kJ/mol. Notably, the results showed that the simplified molecular line entry system (SMILES) with a length of 500–700 has the highest synthetic accessibility (SA), suggesting that molecules with this range of complexity are more easily accessible than those with shorter or longer SMILES. This study demonstrates the potential of ML-assisted design of new organic materials and highlights the importance of considering SA in the design process. The predicted compounds with high ΔH and optimal SA hold promise for various applications, including propellants, explosives, and other high-energy materials.