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

Area of Science:

• Computational Chemistry
• Machine Learning
• Quantum Chemistry

Background:

• Accurate wavefunction methods (e.g., CCSD(T)) provide high precision for small molecules but are computationally expensive.
• Density Functional Theory (DFT) is practical for large systems but often lacks quantitative accuracy for electronic changes.
• Predicting electron affinities (EAs) accurately is crucial for understanding molecular properties.

Purpose of the Study:

• To develop a machine learning model for accurate and efficient prediction of electron affinities.
• To overcome the limitations of computational cost associated with wavefunction methods and accuracy issues with DFT.
• To integrate physics-based and quantum chemistry features for improved predictive power.

Main Methods:

• Developed a machine learning model using an XGBoost Regressor (XGBR) within a ΔML framework.
• Incorporated physics-based structural features (RDKit, SMARTS) and quantum chemistry-based electronic features (Mulliken charge analysis).
• Enhanced the model (ΔML+) by embedding QM-based electronic features for improved accuracy.

Main Results:

• The ΔML+ model achieved a mean absolute error of 0.03 eV compared to G4MP2 values, exceeding chemical accuracy.
• The model demonstrated reduced dependence on the underlying DFT functional.
• Analysis of vertical and adiabatic EAs highlighted the importance of geometry relaxation.

Conclusions:

• The developed ΔML+ model provides an efficient, transferable, and accurate method for predicting electron affinities.
• This approach surpasses conventional chemical accuracy targets and offers a next-generation computational protocol.
• The findings pave the way for overcoming limitations of standalone DFT in predicting molecular properties.