Machine learning approaches to XANES spectra for quantitative 3D structural determination: The case of CO2 adsorption on CPO-27-Ni MOF

Abstract

In this work we have applied machine learning methods (Extra Trees, Ridge Regression and Neural Networks) to predict structural parameters of the system based on its XANES spectrum. We used two ML approaches: direct one, i.e. when ML model is trained to predict the structural parameters directly from the XANES spectrum and inverse one when ML model is used to approximate spectrum as a function of structural parameters. We show the applicability of several ML approaches to predict the geometry of CO2 molecule adsorbed on Ni2+ surface sites hosted in the channels of CPO-27-Ni metal-organic framework. Quantitative fitting is based on difference XANES spectra and we discuss advantages and disadvantages of the two ML approaches and critically examine the overfitting phenomenon, caused by systematic differences of experimental data and learning dataset.