Interpretable Deep Learning Model for Analyzing the Relationship between the Electronic Structure and Chemisorption...

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Machine learning (ML) models offer high predictive performance for material properties but often lack interpretability due to their complexity.
• Developing interpretable ML models is crucial for advancing materials science and enabling universal applications.

Purpose of the Study:

• To develop a deep learning model that extracts human-understandable knowledge from material properties while maintaining high predictive performance.
• To interpret the correlation between a material's surface electronic density of states (DOSs) and its chemisorption property.

Main Methods:

• Developed a deep learning model incorporating an attention mechanism to identify key features in DOSs relevant for predicting adsorption energies.
• Utilized the model to analyze the relationship between electronic structure and chemisorption behavior in materials.

Main Results:

• The deep learning model successfully identified important regions within the DOSs that correlate with chemisorption properties.
• The model autonomously reconstructed the established d-band center theory without prior explicit knowledge, demonstrating its ability to derive scientific principles.
• Achieved high predictive performance in modeling adsorption energies.

Conclusions:

• Complex machine learning models can yield interpretable insights into material behavior.
• The developed attention-based deep learning approach facilitates the extraction of fundamental scientific knowledge, such as the d-band center theory, from data.
• This work paves the way for more transparent and universally applicable ML tools in materials science.