Interpretable, calibrated neural networks for analysis and understanding of inelastic neutron scattering data

Abstract

Deep neural networks provide flexible frameworks for learning data representations and functions relating data to other properties and are often claimed to achieve 'super-human' performance in inferring relationships between input data and desired property. In the context of inelastic neutron scattering experiments, however, as in many other scientific scenarios, a number of issues arise: (i) scarcity of labelled experimental data, (ii) lack of uncertainty quantification on results, and (iii) lack of interpretability of the deep neural networks. In this work we examine approaches to all three issues. We use simulated data to train a deep neural network to distinguish between two possible magnetic exchange models of a half-doped manganite. We apply the recently developed deterministic uncertainty quantification method to provide error estimates for the classification, demonstrating in the process how important realistic representations of instrument resolution in the training data are for reliable estimates on experimental data. Finally we use class activation maps to determine which regions of the spectra are most important for the final classification result reached by the network.

Interpretable and Explainable Machine Learning for Materials Science and Chemistry

“…b) CAM in action -a CNN was trained to classify magnetic Hamiltonians based on inelastic neutron scattering spectra (upper) and highlight the regions of energy transfer in Q-space that are important for making distinctions using a CAM (lower). The regions identified by the CNN/CAM match with the regions that a trained physicist identifies, but in a fraction of the time[30].…”

“…SHAP analysis has been used to understand ML models that predict the efficacy of organic capping layers for increasing stability of halide perovskites solar cells, highlighting the importance of low numbers of hydrogen bond donors and small topological polar surface ares [23]. Salience methods have been used to identify the regions in 3D neutron spectroscopy signals that are most important for deciding the magnetic structure in a double perovskite, these regions are found to match with the regions identified by a trained physicist, but are found in a fraction of the time [30]. Salience methods were also used to identify the regions responsible for misclassifications in an X-ray diffraction analysis deep neural networks, allowing human intervention where the model is likely to perform poorly [52].…”