Interpreting surface adsorption with band data: a machine learning perspective on quantum orientation

A band-based approach with quantum mechanical data is successful in bridging quantum data and practical applications. Quantum mechanical intuitions of a band-based model (e.g. the d-band model) can be related to the adsorption states being reshaped by a band in an atom-like pattern. However, the incompleteness hinder the extension of the band-based model for broader and more accurate applications. Herein, in this work, a systematical extension of the band-based model is introduced with a complete decomposition of bonding/anti-bonding/non-bonding in a consistent and parameter-free way. The Bonding Decomposition with limited aid from machine learning algorithms, possesses enough explainability power which allow people to evaluate hybridization, repulsive orthogonalization and nonbonding occupations. With the new approach, the adsorption energy can be related to the band-center-like single indicator, or be better related to up to 5 better indicators including a high occupation center, a high occupation width, a nonbonding component in high occupation, a bonding indicator and an anti-bonding indicator in low occupation. As the 5 indicators cover the essence of the band center evaluations, the adsorption energies can simply be modeled with a linear model which split “cans” from “cannots” out of the pure Bonding Decomposition. The Bonding Decomposition is applied to metal d-band and perovskite p-band data. After the normally distributed residuals are considered as “random errors”, the Bonding Decomposition may be good for medium adhesion strength. The algorithm displaces why the band center for metal d-band data is already an effective approach and why perovskite p-band data must be handled with the nonbonding contributions considered. Graphical Abstract