Abstract

X-ray Powder Diffraction (XRPD) is a powerful method in material sciences that gives insights into the atomical and crystallographic structure of a material, revealing information into the material's properties and suitability for industrial and scientific application. In geology, XRPD analysis is frequently leveraged to identify and quantify the present mineral phases in an unknown mixture. Despite it's widespread use, interpreting XRPD patterns requires highly-specialized knowledge, making the analysis largely dependent upon the background experience of the analyst. To assist experts, computational methods have been developed over the years. Some of these techniques involve fitting diffraction patterns using pseudo-Voigt functions, which approximate peak shapes through iterative least-squares optimization. However, for complex XRPD patterns, this process can be time-intensive. To address this weakness, we propose an algorithm that utilizes Gradient Boosted Regressors to rapidly generate pattern profiles. Our method achieves a sevenfold speed improvement over traditional approaches while maintaining comparable accuracy in the generated profiles. As computational tools have advanced, several software packages have been developed to automate XRPD analysis. Yet, a major pitfall of these tools is their inability to incorporate domain-specific geological knowledge to their predictions, often resulting in the identification of highly improbable mineral combinations. This disconnect highlights a critical gap between automated and expert-level analysis. To bridge this gap, we explore the use of deep learning for mineral phase identification. We develop a convolutional neural network trained on simulated XRPD patterns to detect the presence of 66 different mineral phases. The model is evaluated on a small test set comprised of 37 empirical patterns from synthetic mineral mixtures, achieving a precision and recall of approximately 70%. By integrating domain-specific constraints into the simulated data, the model overcomes weaknesses of existing automated analysis software by incorporating domain-specific knowledge into it's predictions. This approach demonstrates the potential of data-driven models to augment expert analysis, pioneering the way for more efficient mineral characterization workflows.

Degree

College and Department

Computational, Mathematical, and Physical Sciences; Mathematics

Rights

https://lib.byu.edu/about/copyright/