In the development of new materials, determining material structure is often a major undertaking. Every new material must undergo thorough and detailed structural characterization before its properties can be correlated with performance—yet this step frequently slows down the pace of research and development. Can this process be made faster and more efficient?

Electron backscatter diffraction (EBSD), a technique developed more than 30 years ago and now widely integrated into scanning electron microscopy (SEM), is a critical tool for identifying crystal phases and determining crystallographic orientations. EBSD has further advanced applications in texture analysis and residual stress detection. Scientists can even reconstruct three-dimensional EBSD data from multiple two-dimensional datasets to analyze microstructural features such as grain boundaries. However, traditional orientation-indexing methods have long faced a trade-off between speed and accuracy: while the Hough transform is computationally fast, its accuracy is limited; dictionary indexing and full pattern matching improve precision but are extremely time-consuming and require substantial storage and computational resources.

Assistant Professor Shao-Pu Tsai of the Department of Materials Science and Engineering at National Taiwan University, in collaboration with Dr. Bo-Yan Tung of the University of Cambridge, led a research team in developing a new method called Latice. By leveraging artificial intelligence to automatically learn the physical characteristics of crystals from data, the method successfully overcomes these longstanding bottlenecks, enabling faster and more efficient crystal analysis. Using a variational autoencoder (VAE) to process EBSD diffraction patterns, the team achieved three major breakthroughs: (1) an indexing speed 7.5 times faster than conventional methods; (2) 99.9% data compression, dramatically reducing storage requirements; and (3) the ability to capture crystal rotational symmetry, demonstrating an AI model that “understands physics.”

This three-year research project was primarily carried out by Yu-Chun Liu, who began the work as a senior undergraduate student in the Department of Materials Science and Engineering and later continued as a master’s student in the graduate institute. The team acknowledges funding support from the National Science and Technology Council, the NTU Office of Research and Development, and Walsin Lihwa Corporation. The research成果 has been successfully applied to crystallographic orientation analysis of fully recrystallized 316 stainless steel and has been published in the international journal Cell Reports Physical Science. Looking ahead, the team will continue to advance physics-informed artificial intelligence research and further expand its applications in materials science.