Uncertainty quantification of spectral predictions using deep neural networks

Area of Science:

• Computational materials science
• Spectroscopy
• Machine learning applications

Background:

• Deep neural networks (DNNs) are increasingly used for predicting complex material properties.
• Accurate uncertainty quantification is crucial for reliable DNN predictions in scientific applications.
• Transition metal K-edge X-ray absorption near-edge structure (XANES) spectroscopy provides insights into electronic structure.

Purpose of the Study:

• To evaluate the effectiveness of uncertainty quantification (UQ) methods for DNN-based XANES spectral predictions.
• To compare deep ensembles and bootstrap resampling for UQ in this context.
• To assess the performance of a multi-layer perceptron (MLP) model with UQ for transition metal XANES.

Main Methods:

• Implementation of deep ensembles and bootstrap resampling as UQ techniques.
• Development and application of a multi-layer perceptron (MLP) model for predicting XANES spectra.
• Validation using held-out data for nine first-row transition metal K-edge XANES spectra.

Main Results:

• Bootstrap resampling demonstrated a high degree of accuracy in uncertainty assessment.
• Over 90% of predicted spectral intensities fell within ±3 standard deviations (σ) of the true values for held-out data.
• The MLP model combined with bootstrap resampling provided reliable uncertainty estimates for XANES spectra.

Conclusions:

• Bootstrap resampling is a viable and accurate method for quantifying uncertainty in DNN predictions of transition metal XANES spectra.
• The developed MLP model with bootstrap UQ can be a valuable tool for materials science research.
• This work highlights the importance of UQ for trustworthy AI in scientific discovery.