Abstract by Rachel Drapeau
Reducing user error in QXRD analysis using machine learning
Quantitative x-ray powder diffraction (QXRD) analysis is used to determine the phases contained in rock samples. However, the accurate identification of phases in a sample is largely dependent on the expertise and geologic knowledge of the analyst. To reduce this user error, we plan to incorporate supervised machine learning (SML) algorithms into RockJockML - a MATLAB based program for full-pattern summation QXRD analysis - to mimic the decisions expert analysts make. The data set will be created by using GSAS-II software to generate synthetic XRD patterns for samples by combining known crystal structure information and a calculated probability distribution of Rietveld parameters for each mineral in a sample. The models produced by the SML will then be compared to current progressive optimization techniques using the validation data set to determine which SML model performs best.