Automated EXAFS and Nanoindentation Analysis using Artificial Intelligence in Addressing Reproducibility Challenges

We have developed a novel, artificial intelligence-based methodology that can be utilized to reliably analyze experimental results from Extended X-ray Absorption Fine Structure (EXAFS) measurements and other spectral analyses. This development will help to address (1) the reproducibility problems that slow research progress and inhibit effective technology transfer and manufacturing innovation in these scientific disciplines and (2) ultimately a challenge in obtaining the best model fitting parameters from the large and growing quantities of real-time materials characterization data produced by modern instruments, with minimal human intervention.<br/>The existing EXAFS analysis approach relies on a human analyst to suggest a potential set of chemical compounds in the form of feff.inp input files that may be present, which can cause reproducibility issues. The situation is made even worse with advances in instrumentation that have been enabled to produce data 2-3 orders of magnitude more rapidly than it can be analyzed.<br/>Instead, we applied a machine learning approach to the analysis of EXAFS spectroscopy measurements collected using a synchrotron radiation facility. Specifically, for the first time, we developed a genetic algorithm (GA) for fitting the measured spectra to extract the relevant structural parameters. The algorithm attempts to determine the best structural paths from these compounds that are present in the experimental measurement. It starts with a population consisting of a number of temporary fittings and looks for the primary EXAFS path contributors from the potential compounds. It calculates goodness of fit value that can be used to identify the chemical moieties present. To improve the accuracy and efficiency of finding an optimized final solution, evolutionary-inspired operators (e.g., crossover, mutator) are applied to each solution throughout subsequent generations. The advantage of using GA is that it can explore large and complex parameter spaces for model solutions and does not require computation of the derivatives of the functional objectives, nor assumptions of continuity and convexity for the objective functions and constraints.<br/>The analysis package is called EXAFS Neo and is open-source and written in Python. It requires using Larch and Feff to calculate the initial EXAFS paths. We have recently extended the code to make use of Feff8.5lite so it can calculate the paths needed for populating the analysis from within the EXAFS Neo package. We have also expanded the Neo to fitting Nanoindentation, the analysis of core-level photoemission, and spectral fitting in other fields like X-ray astronomical data and published the analysis results.