XElemNet: towards explainable AI for deep neural networks in materials science

Area of Science:

• Materials Science
• Artificial Intelligence
• Computational Chemistry

Background:

• Deep learning accelerates materials discovery but often functions as a "black-box," raising interpretability concerns.
• ElemNet, a deep neural network, predicts formation energy from elemental composition, showcasing AI's potential in materials science.

Purpose of the Study:

• To enhance the interpretability and reliability of the ElemNet deep learning model.
• To apply explainable artificial intelligence (XAI) techniques for post-hoc analysis and model transparency.

Main Methods:

• Utilized XAI techniques to analyze the ElemNet model.
• Conducted experiments using artificial binary datasets.
• Performed feature importance analysis on ElemNet's predictions.

Main Results:

• ElemNet effectively predicts convex hulls for element-pair systems across periodic table groups.
• The model demonstrates capability in discerning elemental interactions.
• Feature importance analysis revealed alignment with key chemical properties like reactivity and electronegativity.

Conclusions:

• XElemNet provides crucial insights into ElemNet's strengths and limitations.
• This work offers a pathway for explaining other deep learning models in materials science.
• Enhancing AI interpretability is vital for its reliable application in scientific discovery.