Machine learning approach for predicting coordination numbers from EXAFS spectra

BackgroundX-ray Absorption Fine Structure (XAFS) is a vital technique for structural analysis, but the complexity of XAFS spectra often requires interpretation by experienced researchers, which can still lead to inaccuracies.PurposeThis study aims to develop machine learning models for predicting coordination numbers directly from Extended X-ray Absorption Fine Structure (EXAFS) spectra of fourth-period transition metal elements.MethodsFirstly, a dataset of 13 374 valid EXAFS spectra was collected from the Materials Project database, covering diverse coordination environments of fourth-period transition metals. Secondly, the EXAFS data were transformed from energy space to R-space using Fourier transformation, with intensity values extracted at 0.003 nm intervals as features. Thirdly, three machine learning models, i.e., neural networks, bagging decision trees, and random forests, were trained and evaluated using five-fold cross-validation to predict the coordination numbers of absorbing atoms. Finally, feature importance analysis based on the Gini index was performed to identify key spectral regions contributing to predictions.ResultsPrediction results show that three models achieve an average prediction accuracy of approximately 70%, with the neural network reaching up to 81.74% for vanadium. Feature importance analysis reveals that data points at RConclusionsResults of this study demonstrate that machine learning models, particularly random forests with Gini-based feature analysis, provide interpretable and reliable tools for EXAFS data analysis. The achieved ~70% average accuracy across diverse transition metal systems indicates strong generalizability, while the identification of critical spectral features (R<0.3 nm) validates the physical basis of model predictions and enhances the efficiency of automated structural characterization.