Pair distribution function analysis for oxide defect identification through feature extraction and supervised learning

TL;DR

This paper develops a machine learning approach using feature extraction from pair distribution functions to identify and quantify defects in TiO₂, combining physics-based models with neural networks for improved material characterization.

Contribution

It introduces a novel pipeline integrating physical feature extraction and supervised learning to accurately predict defect types and concentrations from experimental PDFs.

Findings

Autoencoder with physics-based initialization yields highest accuracy.

The method outperforms brute-force predictions on experimental data.

Features extracted effectively correlate with defect concentrations.

Abstract

Feature extraction and a neural network model are applied to predict the defect types and concentrations in experimental TiO$_2$ samples. A dataset of TiO$_2$ structures with vacancies and interstitials of oxygen and titanium is built and the structures are relaxed using energy minimization. The features of the calculated pair distribution functions (PDFs) of these defected structures are extracted using linear methods (principal component analysis, non-negative matrix factorization) and non-linear methods (autoencoder, convolutional neural network). The extracted features are used as the inputs to a neural network that maps the feature weights to the concentration of each defect type. The performance of this machine learning pipeline is validated by predicting the defect concentrations based on experimentally-measured TiO$_2$ PDFs and comparing the results to brute-force predictions. A physics-based initialization of the autoencoder has the highest accuracy in predicting the defect concentrations. This model incorporates physical interpretability and predictability of material properties, enabling a more efficient material characterization process with scattering data.