QSPR models for sublimation enthalpy of energetic compounds

The sublimation enthalpy of energetic compounds is often predicted using quantum chemistry (QC) based quantitative structure-property relationship (QC-QSPR), which is accurate but requires high CPU cost. A feasible alternative is machine learning (ML), but it lacks applicability for energetic molecules, due to the limited experimental data thereof. A new data set for sublimation enthalpy is established, by extending a commonly used one with that of energetic organic compounds collected from literatures. Four topological descriptors are proposed to construct QSPRs, which exhibit higher accuracy than the QC-based ones, and are used to build ML based QSPRs with the four algorithms individually. The Extreme Gradient Boosting (XGBoost) model exhibits the highest accuracy, with the mean absolute error of 2.7 kcal/mol, followed by the Particle Swarm Optimization (PSO) one. Still, the PSO model is more portable and recommendable, because it is fully interpretable. The PSO model can accurately predict sublimation enthalpy with negligible CPU time cost, and is expected to be used to find novel energetic molecules by further predicting detonation properties.

Automated summary

This study establishes a new dataset and uses machine learning methods to predict the sublimation enthalpy of energetic compounds. The results show that QSPR models constructed with four topological descriptors are more accurate than QC-based models, with the Particle Swarm Optimization model being a well-performing and interpretable model.