Presentation Information
[3F09]Bayesian Inference for Comprehensive and Rapid Crystalline Phase Identification from X-ray diffraction
*Kenji Nagata1, Ryo Murakami1 (1. National Institute for Materials Science)
In this talk, we introduce our proposed framework of Bayesian estimation that identifies optimal combinations of crystalline phases by analyzing the entire diffraction profile. Our approach calculates the posterior probabilities of all possible phase combinations through an approximate exhaustive search, thereby incorporating both the shapes of all diffraction peaks and their possible overlaps due to multiple phases. While this method significantly improves reliability, it requires several hours to complete a single analysis due to its computational complexity. We also developed a new method to overcome this limitation by introducing a fast, practical Bayesian phase identification method that delivers results within seconds. We integrate variational sparse inference with GPU acceleration to drastically reduce computation time. Our method achieves phase identification in under 10 seconds, even when evaluating 250 candidate phase combinations. Moreover, the phases identified by our approach are consistent with those reported in previous studies using more computationally intensive, high-precision algorithms.