Explainable artificial intelligence for materials discovery: application to catalysts for the HER and ORR

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Machine learning (ML) accelerates materials property prediction and virtual design.
• Current ML models often function as "black boxes," limiting insights into material behavior.
• Explainable artificial intelligence (XAI) is crucial for understanding ML predictions and uncovering new scientific knowledge.

Purpose of the Study:

• To develop a novel, explainable ML strategy for materials design.
• To integrate counterfactual explanations into ML pipelines for materials discovery.
• To provide insights into structure-property relationships beyond simple prediction accuracy.

Main Methods:

• Utilized counterfactual explanations for ML-driven materials design.
• Employed explainable artificial intelligence (XAI) tools within the ML workflow.
• Validated discovered materials using density functional theory (DFT) calculations.
• Analyzed relationships between material features and target properties using explanations.

Main Results:

• Successfully identified novel materials with properties close to design targets.
• Validated the performance of these materials using rigorous DFT calculations.
• Unveiled subtle correlations between material features and desired properties through counterfactual analysis.
• Demonstrated the capability of XAI to provide deeper chemical and physical insights.

Conclusions:

• The proposed counterfactual explanation strategy offers a powerful, explainable alternative to traditional materials design methods.
• Integrating XAI into ML for materials science enhances discovery by providing actionable insights.
• This approach facilitates a deeper understanding of materials, accelerating innovation in areas like catalysis.