Hybrid Descriptor-Based Machine Learning and COSMO-RS Modeling for Accurate Flash Point Prediction of Organic Chemicals

The current research on flash point prediction models is still limited by the applicability scale and model accuracy. To address this issue, three different machine learning (ML) models, namely, the artificial neural network (ANN), eXtreme Gradient Boosting (XGBoost), and Gaussian Process Regression (GPR) models, were constructed and compared by applying either COnductor-like Screening MOdel (COSMO)-derived features or the hybrid molecular structure descriptors (COSMO-MSD). The results demonstrated that mixed features provide more comprehensive information than the single-COSMO feature. Among the six models evaluated, the GPR-COSMO-MSD model achieved the highest accuracy ( R 2 = 0.932). The Shapley additive explanation (SHAP) analysis further identified the top 20 influential input features in XGBoost-COSMO-MSD and GPR-COSMO-MSD models, which indicates a totally discrepant feature significance yet identifies common impactful features, such as TPSA, LabuteASA, and BertzCT. These findings from the SHAP analysis can provide valuable insights for selecting the key factors for accurate flash point predictions.