What does AI look at when selecting materials?
A new method reveals how AI reaches its predictions, enabling more efficient discovery of optical materials
What the research is about
The brilliant colors of gemstones, the vivid displays of smartphones, and the performance of solar cells may seem unrelated at first glance. In fact, they all depend on how materials respond to light.
To develop such materials, scientists need to understand which combinations and arrangements of atoms produce the properties they want. But the number of possible combinations is enormous, making it impossible to test every candidate experimentally. Finding a new material can take years—or even decades.
To speed up this process, researchers have increasingly turned to artificial intelligence (AI). Many AI models have already been developed to predict single numerical properties of materials, such as hardness or melting point. However, relatively few studies have focused on predicting more complex data, such as how a material responds to different colors (wavelengths) of light. Even when these models make accurate predictions, they rarely explain how they arrived at their answers.
A research team led by Associate Professor Akira Takahashi at Institute of Science Tokyo (Science Tokyo), developed a new analytical method that not only trained AI using data from 2,681 inorganic materials but also revealed how the AI made its predictions. Their approach makes it possible to understand which clues the AI uses when predicting the complex optical properties of materials.
Why this matters
By examining the inner workings of the AI with their new analytical method, the researchers uncovered the clues the AI relied on when making its predictions. The AI automatically grouped together materials with similar optical responses and similar predictive patterns such as which elements are important and how specific atomic arrangements give rise to desired properties.
Even more surprisingly, the AI had never been explicitly taught advanced chemical concepts. Nevertheless, during training it learned to organize materials in ways that closely matched the chemical principles already understood by scientists.
This study is therefore about more than building an accurate prediction model. It represents a major step forward because it enables researchers to understand how AI reaches its predictions, making AI a more transparent and useful tool for scientific discovery.
What’s next
This research has revealed which combinations of atoms and crystal structures the AI considers important when predicting material properties. These insights could help researchers search more efficiently for new optical materials, including high-performance solar cells, light sensors, environmentally friendly coatings, and coloring materials.
The method is not limited to optical properties. It could also be applied to many other types of complex data that change with factors such as time, temperature, or pressure. As a result, it has the potential to accelerate discoveries not only in materials science but also in many other research fields that rely on analyzing high-dimensional data.
Comment from the researcher
AI is becoming increasingly important in materials science, and its ability to predict material properties continues to improve in both speed and accuracy. However, AI is not perfect. Its performance can decline when dealing with materials outside the range of its training data, and researchers must still decide which materials should be studied and what data should be used for learning. Scientific theories and ideas therefore remain essential. Rather than using AI simply as a prediction tool, our approach helps interpret how AI reaches its conclusions. We hope this will lead to new scientific hypotheses and design ideas, further accelerating materials science and the discovery of new materials.
(Akira Takahashi, Associate Professor, Materials and Structures Laboratory, Institute of Integrated Research, Institute of Science Tokyo)
Dive deeper
Explore more research in Science for All
Science for All showcases cutting-edge research at Science Tokyo and highlights the ideas, people, and possibilities shaping the future.
Contact
Research Support Service Desk