Deep learning for visualization and novelty detection in large X-ray diffraction datasets

Lars Banko,Phillip M Maffettone,Dennis Naujoks,Daniel Olds,Alfred Ludwig

doi:10.1038/s41524-021-00575-9

Lars Banko, Phillip M Maffettone + Show 3 more

Open Access

https://doi.org/10.1038/s41524-021-00575-9

Copy DOI

Abstract

We apply variational autoencoders (VAE) to X-ray diffraction (XRD) data analysis on both simulated and experimental thin-film data. We show that crystal structure representations learned by a VAE reveal latent information, such as the structural similarity of textured diffraction patterns. While other artificial intelligence (AI) agents are effective at classifying XRD data into known phases, a similarly conditioned VAE is uniquely effective at knowing what it doesn’t know: it can rapidly identify data outside the distribution it was trained on, such as novel phases and mixtures. These capabilities demonstrate that a VAE is a valuable AI agent for aiding materials discovery and understanding XRD measurements both ‘on-the-fly’ and during post hoc analysis.

Highlights

Innovations in high-throughput and autonomous experimentation[1,2,3,4] are exceedingly increasing the acquisition rate of data, in the case of X-ray diffraction (XRD)
Recent progress has been made in using artificial intelligence (AI) for unsupervised XRD dataset decomposition[6,7], crystal structure classification[8,9,10,11,12,13,14,15,16,17,18,19,20], and integrating the latter with autonomous experimentation[21]
We develop a dynamic visualization tool for experimental XRD patterns and the variational autoencoders (VAE) latent space. While this offers a correlation with structural classification, we show that visualizing the latent space with respect to the reconstruction error of the VAE allows for novelty detection during an experiment

Summary

INTRODUCTION

Innovations in high-throughput and autonomous experimentation[1,2,3,4] are exceedingly increasing the acquisition rate of data, in the case of XRD. Experiment specific models have a more relevant, yet more narrow, distribution or training data, and struggle when encountering data outside of this distribution It is the domain knowledge and prior information (e.g. simulated X-ray diffractograms of expected phases from crystallographic database entries that encompass the breadth of possible experimental non-idealities) that makes these AI agents so successful[18]. We use a variational autoencoder (VAE) (Fig. 1a) trained on a synthetic dataset as a prior[18,23] to solve commonly occurring visualization and novelty detection challenges in XRD analysis This same synthetic dataset can be used to train a state-of-the-art classification model in tandem[18,24]. While this offers a correlation with structural classification, we show that visualizing the latent space with respect to the reconstruction error of the VAE allows for novelty detection during an experiment

RESULTS AND DISCUSSION

METHODS

CODE AVAILABILITY