Interpretable AI in materials discovery: Uncovering how models make predictions

Researchers from the Institute of Science Tokyo have developed a new method to interpret AI models used in materials discovery by analyzing their learned features.
The method combines a graph neural network with hierarchical clustering to extract key features linking crystal structure to optical spectra.
It groups materials with similar structural and spectral characteristics, revealing patterns that can guide efficient material design.
The approach, using an ALIGNN and hierarchical clustering, enabled classification of materials based on both structural features and spectral shapes.
The interpretability of these AI models enhances understanding of how atomic arrangements influence material properties, beyond just optical spectra.
The team, led by Assistant Professor Akira Takahashi, demonstrated the potential of the method with data on metal oxides, chalcogenides, and related compounds.
The method offers insights into how material structures impact various properties under different conditions, not limited to optical spectra.
The study aims to provide useful physical and chemical insights for materials design and will be published in the journal Advanced Intelligent Discovery.

You may have missed