Explainable AI-based physical theory for advanced materials design
Microscopic materials analysis is essential to achieve desirable performance in next-generation nanoelectronic devices, such as low power consumption and high speeds. However, the magnetic materials involved in such devices ...
Traditionally, researchers have performed a visual analysis of the microscopic image data. However, this often makes the interpretation of such data qualitative and highly subjective. What is lacking is a causal analysis of the mechanisms underlying the complex interactions in nanoscale magnetic materials.
In a recent breakthrough published in Scientific Reports, a team of researchers led by Prof. Masato Kotsugi from Tokyo University of Science, Japan succeeded in automating the interpretation of the microscopic image data. They achieved this using an "extended Landau free energy model" that they developed using a combination of topology, data science, and free energy.
The model illustrated the physical mechanism as well as the critical location of the magnetic effect, and proposed an optimal structure for a nanodevice. The model used physics-based features to draw energy landscapes in the information space, which could be applied to understand the complex interactions at the nanoscales in a wide variety of materials.
"Conventional analysis are based on a visual inspection of microscope images, and the relationships with the material function are expressed only qualitatively, which is a major bottleneck for material design. Our extended Landau free energy model enables us to identify the physical origin and location of the complex phenomena within these materials. This approach overcomes the explainability problem faced by deep learning, which in a way amounts to reinventing new physical laws," Prof. Kotsugi explains.
An image depicting the extended Landau free energy model developed by a research team from Tokyo University of Science, which enables a causal analysis of the magnetization reversal in nanomagnets. Through this model, the team could visualize magnetic domain images effectively and were successful in the inverse designing of nanostructures with low energy requirements. Credit: Kotsugi Laboratory from Tokyo University of Science, Japan.
Scatterplot of the dimensionality reduction results of principle component analysis. Color represents the total energy. The relationship between magnetic domain and total energy is connected in the explainable feature space. Credit: Masato Kotsugi from Tokyo University of Science, Japan.
Scientists from TUS have succeeded in visualizing slight changes in microscopic images and understanding the mechanisms that have been difficult to analyze visually. Furthermore, they have succeeded in inverse designing nanostructures with low energy consumption. Credit: Masato Kotsugi from Tokyo University of Science, Japan.