A Machine-Learning Approach to Quantitative Phase Analysis of Whole Rocks and Sediments

点击翻译按钮获取中文摘要

Studies of the chemistry of natural rocks and sediments usually require quantification of how much of different minerals and related materials are present. This is true for academic studies, as well as those by environmental consulting, petroleum extraction, and mining companies. This project is meant to improve the quality of a standard method of making these measurements, powder X-ray diffraction. The main problem with using powder X-ray diffraction for such a purpose is that when multiple minerals are present, the analyst needs to supply a rather limited list of possible minerals for which to test. To do this successfully, the analyst must have a certain amount of background knowledge about which minerals are common or rare, and which tend to occur together in the same geologic environments. Consequently, novice analysts are notorious for producing implausible, or even strange, results. This project will address this problem by using artificial intelligence techniques to make the process of choosing minerals to include in an analysis more automated, allowing novice analysts to produce more reliable results.The project will test the hypothesis that computer algorithms generated via machine learning techniques can enhance a popular powder X-ray diffraction analysis technique to deliver credible quantitative phase analyses in an automated, or semi-automated, manner. The starting framework is a program (RockJockML) that fits sample diffraction patterns with experimental patterns in a library that includes ~150 minerals. A large synthetic data set of powder X-ray diffraction patterns and X-ray fluorescence analyses will be generated using crystal structure data from the American Mineralogist Crystal Structure Database for mineral specimens that are similar to those in RockJockML's standards library. The individual samples in the synthetic data set will be modeled as mixtures of 1-15 minerals found together at randomly chosen locations included in the Mineral Evolution Database (MED), so that the data set will mimic known mineral prevalence and co-occurrence patterns. Various machine learning algorithms will be fed the training sets to produce models that best predict which phases are included in each sample. The best models will then be incorporated into RockJockML, and versions of the software will be produced in MATLAB and Python. The investigators will also produce a database tool for adding phases to the RockJockML database.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.