Machine learning-based XANES analysis for predicting the local structure and valence in amorphous silicon suboxides.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Silicon suboxide (SiO_x) exhibits tunable properties based on its composition, making it industrially relevant.
• Accurately correlating the atomistic structure of SiO_x with its properties is challenging with traditional methods.
• Understanding SiO_x requires quantitative analysis of its electronic valence and local atomic structure.

Purpose of the Study:

• To develop a machine learning model for predicting silicon atom valence state and Si-O radial distribution function from X-ray absorption near-edge structure (XANES) spectra.
• To establish a robust and experimentally viable framework for characterizing amorphous silicon suboxide materials.

Main Methods:

• Generated nine amorphous SiO_x networks using molecular dynamics simulations.
• Calculated Si K-edge XANES spectra using first-principles calculations.
• Trained a deep neural network on the generated XANES spectra dataset.

Main Results:

• The deep neural network model accurately predicts local silicon valence state and Si-O radial distribution function from XANES spectra.
• Specific spectral regions were identified as crucial for valence state (near edge) and structural (higher energy) predictions.
• The model demonstrated robustness, maintaining high performance on composition-averaged spectra.

Conclusions:

• The developed machine learning approach enables direct extraction of electronic and structural information from XANES spectra of amorphous materials.
• This method overcomes a significant bottleneck in analyzing complex, multivalent amorphous systems.
• The framework facilitates accelerated development of SiO_x-based functional materials by quantitatively characterizing composition-structure-property relationships.