Interpretable Machine learning based coordination motif identification scheme from X-ray absorption near-edge structure spectroscopy XANES

Abstract

XANES is an important experimental method to probe the local three dimensional geometry and electronic structure of the system. The quantitative analysis of XANES data is very important to obtain the above mentioned structure. Because XANES contains a lot of information and complexity, the quantitative analysis of XANES is also a challenging task. In this paper, the coordination number and stereo coordination motif of the system are analyzed using the random forest, XGBoost, LightGBM algorithm based on the small and medium scale XANES theoretical spectral database. The accuracy of the analytical results is 90% in coordination number aspect and more than 80% in stereo coordination motif, which has practical value. At the same time, we have carried out analysis based on interpretable machine learning. We use the TreeSHAP method to explain the mechanism of our model and of a specific prediction example. The model mechanism is explained from the physical perspective as much as possible, which expands the methodological perspective of machine learning application in XAS data analysis. This method builds machine learning model based on database, which has great generality and high model repeatability. This method can be used to provide the initial structure for on-line system judgment during beamline experiment time, and can also be used as the starting point for subsequent XANES fitting or neural network spectral analysis. This method is based on the small and medium scale database construction model, which can be used as a reference for construction of beamline experiment spectrum database and related data analysis produce. This work has formed a preliminary GUI applet, which will be a submodule of an overall XAS data analysis framework in the future.