What Is AI Looking at When It Selects Materials? (Institute of Science Tokyo)

Explanation

Associate Professor Ryo Takahashi of the Institute of Frontier Materials, Institute of Integrated Research at Institute of Science Tokyo, and colleagues have developed a new analysis method to decipher 'what clues AI relies on when predicting the properties of materials.'

In the study, AI was trained on 2,681 types of inorganic materials, and the substance of its judgments was then analyzed. The results showed that the AI automatically grouped similar materials together and discovered on its own 'important elements' and 'the relationship between atomic arrangement and the target property.'

Previous AI research on materials had mainly focused on predicting 'properties expressible as a single number,' such as hardness or melting point. This study goes further, tackling the more complex prediction of optical properties — 'how a material responds to light of various colors.' What is new is that, on top of this, the AI's judgment itself was made interpretable.

Whether humans can verify not just whether an AI's answer is correct but 'why it answered that way' is a deciding factor in whether AI can be trusted as a tool. If an answer appears correct but the reasoning cannot be explained, there is no way to fix it when it turns out to be wrong.

Paper DOI: 10.1002/aidi.202600007. Note that the specific name of the AI method used is not stated in the publicly available information.