Machine Learning-Based Prediction of Bond Dissociation Energies for Metal-Trifluoromethyl Compounds†

Abstract

This study explores the application of machine learning to predict the bond dissociation energies (BDEs) of metal-trifluoromethyl compounds. We constructed a dataset comprising 2219 metal-trifluoromethyl BDEs using density functional theory (DFT). Through a comparative analysis of various machine learning algorithms and molecular fingerprints, we determined that the XGBoost algorithm, when combined with MACCS and Morgan fingerprints, exhibited superior performance. To further enhance predictive accuracy, we integrated chemical descriptors alongside multiple fingerprints, achieving an R² value of 0.951 on the test set. The model demonstrated excellent generalization capabilities when applied to synthesized metal-trifluoromethyl compounds, highlighting the critical role of chemical descriptors in improving predictive performance. This research not only establishes a robust predictive model for metal-trifluoromethyl BDEs but also underscores the value of incorporating chemical insights into machine learning workflows to enhance the prediction of chemical properties.