Proceedings of MATSUS Spring 2024 Conference (MATSUS24)
Publication date: 18th December 2023
Machine learning models that predict properties from a material's composition have, in some cases, proved surprisingly competitive with models that learn a relationship between properties and structure. Electronegativity equalization (EEq) is an older and simpler method than modern machine learning, but can illuminate the relative performance of composition-based and structure-based models. With parameters corresponding to the electronegativity and hardness of each element, EEq provides a model of the partial charge of each atom, based only on the element of the atom and the chemical formula of the molecule. The structure of the molecule can be exploited by including additional interaction terms, which have proved difficult to accurately model. A better understanding of the EEq interaction term may provide a model for understanding the effect of neglecting structure in data-based extrapolation.
We have performed DFT of molecules with two heavy atoms in varying electric fields. We have fit EEq parameters from this dataset using a computational method based on linear regression. In addition, we have estimated the same parameters from energy decomposition analysis of individual molecules, providing molecule-specific parameters and testing the transferability of atoms assumed by EEq. These results also show some promise in illuminating the EEq interaction term.