Physics-informed machine learning model for accurate prediction of electron affinities

Abstract

Accurate wavefunction methods such as CCSD(T) can predict chemical and thermodynamic properties of small molecules with near-experimental precision. However, their steep computational cost limits their use for large systems or extensive datasets. In contrast, density functional theory (DFT) is faster and more practical for large molecules but often fails to accurately capture electronic changes with quantitative accuracy. To address these limitations, we have developed a machine learning model for the prediction of electron affinities with physics-based structural features (RDKit and SMARTS) and quantum chemistry-based electronic features (Mulliken charge analysis). These features, which capture both local bonding motifs and global electronic context, are used as input descriptors to an XGBR (eXtreme Gradient Boosting Regressor) model in a ΔML framework. By embedding QM-based electronic features to the ΔML model, our enhanced ΔML+ model attains a mean absolute error of 0.03 eV with respect to G4MP2 values, surpassing conventional chemical accuracy targets while exhibiting markedly reduced dependence on the underlying DFT functional. By analyzing both vertical and adiabatic EAs within the same framework, we highlight the crucial role of geometry relaxation in predictive modeling. Overall, our approach offers an efficient and transferable route to benchmark accurate electron affinity predictions, pointing toward next generation computational protocols that overcome the limitations of standalone DFT.