R1: Characterization and description of minerals (Joint Session with The Gemmological Society of Japan)

5:45 PM - 6:00 PM JST(8:45 AM - 9:00 AM UTC)

[R1-09] Development of a mineral identification method using SEM-EDX spectra based on deep learning

Ryosei Hirai¹, *Yusuke SETO¹ (1. Osaka Metropolitan University)

Keywords:

Mineral identification,Deep learning

The EDX spectrum varies depending on detector and sample conditions. Therefore, to perform accurate quantitative analysis, it was necessary to prepare many standard materials with guaranteed chemical composition and conduct careful preliminary measurements. Furthermore, the analysis values obtained are merely weight ratios of elements or oxides, so a certain degree of experience is required to identify solid solution minerals with complex chemical compositions. In recent years, deep learning methods have become widespread and are beginning to be used in many fields. On the other hand, EDX spectrum simulation technique has also advanced, making it possible to reproduce high-precision spectra. Therefore, in this study, we developed a new mineral classification method that combines deep learning and EDX spectrum simulation with the aim of further advancing mineral research using EDX analysis.To create a classification model or regression model using deep learning, a large amount of diverse data is required as training data (EDX spectra). However, obtaining such data in actual experimental systems is practically impossible. Therefore, in this study, we created training data from DTSA-II, software developed by NIST, which simulates high-precision EDX spectra. Since the simulation conditions must closely reproduce those of the actual SEM-EDX system, this study comprehensively varied various parameters for standard samples obtained from the actual SEM-EDX system and adopted the parameter set with the highest reproducibility. In addition, by modifying the code to enable multiple simultaneous executions of DTSA-II, we have succeeded in significantly improving the simulation speed. Using the spectra generated by DTSA-II as training data, we constructed a deep neural network model. To evaluate the validity of the proposed method, we divided the olivine solid solution ([Mg,Fe]2SiO4) into 1,000 segments and created a regression model using their simulated EDX spectra. When this regression model was applied to the spectra of peridotite obtained by actual SEM-EDX, we confirmed that the analysis results were accurate to within approximately 1 mol%.