Researchers in Japan have developed an interpretable AI method to explain how AI models make predictions in materials discovery. The method analyses features learned by a trained AI model and uses them to identify relationships between atomic structure and optical spectra.
The study was led by researchers from the Institute of Science Tokyo, in collaboration with Tohoku University. The work is expected to be published in the journal Advanced Intelligent Discovery.
AI is increasingly used in materials research to predict how materials behave based on atomic structure. Such models can accelerate materials discovery and reduce reliance on trial-and-error experimentation, but many operate as black boxes, making it difficult to understand how they arrive at specific predictions.
The researchers addressed this problem by analysing a trained AI model that predicts optical absorption spectra from atomic structural data. They extracted features from the model’s internal layers and clustered materials according to shared structural and spectral characteristics.
The team used an atomistic line graph neural network trained on data from 2,681 metal oxides, chalcogenides, and related compounds. The clustering process classified materials into groups sharing structural characteristics such as elemental composition, atomic coordination, bond lengths, bond angles and similar spectral signatures.
According to the researchers, the model learned meaningful relationships between atomic structure and material properties without being explicitly provided oxidation states or electronic configurations as input. The interpretable AI method could therefore help researchers identify the factors behind desired spectral shapes and support more rational materials design.
The approach could also be applied beyond optical absorption spectra. Researchers said the approach could also help explain how atomic arrangements influence other material properties under varying conditions, such as temperature and pressure, opening new possibilities for designing materials with targeted characteristics.
Why does it matter?
One of the main challenges facing the use of AI in scientific research is explainability. While AI systems can identify patterns and generate accurate predictions, researchers often need to understand the reasoning behind those predictions before they can confidently apply them in experimental settings.
By revealing how AI models connect atomic structures with material properties, interpretable AI could make machine learning a more effective tool for scientific discovery. The approach may help accelerate the development of advanced materials for applications ranging from renewable energy and electronics to sensors and next-generation manufacturing, while improving trust in AI-assisted research.
Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!
