The proposed method combines a graph neural network with hierarchical clustering to extract key features linking crystal structure to optical spectra, and then groups materials with similar structural and spectral characteristics, revealing patterns that can guide materials design. Credit: Institute of Science Tokyo
A method to interpret artificial intelligence (AI) models used in materials discovery by analyzing their learned features has been developed by researchers from Japan. The method extracts key features from an AI model trained on atomic structural data and optical absorption spectra, and then groups materials with similar structural and spectral characteristics. This approach can be extended to reveal how atomic arrangements influence other material properties, paving the way for more efficient materials design.
In recent years, artificial intelligence (AI) has emerged as a powerful tool to predict how materials will behave based on their atomic structure, helping researchers discover new materials faster and reduce reliance on trial-and-error methods. However, many of these models work like “black boxes.” They can make accurate predictions, but they do not explain how those predictions are made. This makes it difficult to understand the relationships between a material’s structure and its properties, limiting how useful these models are for guiding the development of new designs.
Now, in a study to be published in the journal Advanced Intelligent Discovery, researchers from the Institute of Science Tokyo (Science Tokyo), Japan, have developed a method to make these models more interpretable. Their approach works by analyzing a trained AI model and extracting the key features it has learned about how crystal structure relates to optical spectra. Using these features, the researchers then grouped materials that share similar optical spectra and structural characteristics.
The study was led by Assistant Professor Akira Takahashi, Professor Fumiyasu Oba (also a project leader at KISTEC, Japan), and master’s course student Arata Takamatsu (at the time of the research) of the Materials and Structures Laboratory, Science Tokyo, in collaboration with Professor Yu Kumagai of the Institute for Materials Research, Tohoku University, Japan.
“Our proposed classification method allows for an understanding in detail of how AI prediction models make predictions, namely, extracting key factors for desired spectral shapes, and thereby providing useful physical and chemical insights for materials design,” Takahashi says.
Why spectral data are harder
A material’s properties often depend on some parameters and are described using spectral data—for example, optical absorption spectra capturing how light interacts with the material across different wavelengths. Compared with properties represented by a single number, spectral data are far richer and more complex, making them challenging to interpret using conventional AI methods.
The researchers used an atomistic line graph neural network (ALIGNN), an existing graph neural network architecture, trained to predict optical absorption spectra from atomic structure using data from 2,681 metal oxides, chalcogenides and related compounds. From the trained model, they extracted features from its internal layers and applied hierarchical clustering, a method that groups items based on similarity.
This allowed them to classify materials into distinct groups that shared both structural features, such as elemental composition, atomic coordination, bond lengths and bond angles, and similar spectral shapes. Notably, the model learned these patterns from atomic structure alone, without being given oxidation states or electronic configurations as input, indicating that it had internally captured meaningful relationships between structure and properties.
Broader uses for materials design
Optical properties play a key role in many applications. They affect a material’s appearance, which is important for pigments and dyes, and govern how it interacts with light in devices such as solar cells and photodetectors. Understanding which elemental species and structural features shape these spectra is, therefore, key to establishing rational design guidelines for such materials.
Discover the latest in science, tech, and space with over 100,000 subscribers who rely on Phys.org for daily insights.
Sign up for our free newsletter and get updates on breakthroughs,
innovations, and research that matter—daily or weekly.
Furthermore, the approach is not limited to optical spectra: It can be extended to determine how a material’s structure influences its behavior under different conditions, such as temperature or pressure, opening up new possibilities for designing materials with specific and useful properties. As demonstrated here for optical absorption, the approach can be applied to a range of spectral properties, enabling researchers to identify common factors shared by different materials and infer the origins of desired spectral characteristics.
“It has been difficult to interpret what machine learning models have learned about spectral properties. In this work, we developed a general method to extract such insights, which we believe will prove broadly useful for materials research,” Takahashi concludes.
More information
Deep Learning-Based Extraction of Promising Material Groups and Common Features from High-Dimensional Data: A Case of Optical Spectra of Inorganic Crystals, Advanced Intelligent Discovery (2026). DOI: 10.1002/aidi.202600007
Provided by
Institute of Science Tokyo
Who’s behind this story?
Citation:
Interpretable AI in materials discovery: Uncovering how models make predictions (2026, June 14)
retrieved 14 June 2026
from https://phys.org/news/2026-06-ai-materials-discovery-uncovering.html
This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no
part may be reproduced without the written permission. The content is provided for information purposes only.