Autonomous anomaly detection in MeV ultrafast electron diffraction

MeV ultrafast electron diffraction (MUED) is a pump-probe technique to measure dynamic material structure evolution. An ultrashort laser initiates a structure change which is probed by an ultrashort relativistic electron beam. Diffraction patterns are integrated over many shots to beat low signal-to-noise ratio. However, electron beam instabilities from shot to shot disturb the patterns and increase uncertainty. To enhance the accuracy of MUED, anomalous patterns should be detected and removed from datasets with thousands of patterns.

In this work, we developed a machine learning approach to enable autonomous detection of anomalous diffraction patterns. We constructed a convolutional autoencoder model that reconstructs measured patterns of Ta₂NiS₅. We evaluated a one-class support vector machine to detect anomalies based on the distribution of: 1. the feature vectors, 2. the reconstruction errors, and implemented: 3. dimensionality reduction of the reconstruction error by principal component analysis or restricted Boltzmann machine. This hybrid structure allows unsupervised anomaly detection constituting a powerful tool to enhance the accuracy of MUED.