Neural network-based anomaly detection for high-resolution X-ray spectroscopy

We propose an anomaly detection technique for high-resolution X-ray spectroscopy. The method is based on the neural network architecture variational autoencoder, and requires only {\it normal} samples for training. We implement the network using Python taking account of the effect of Poisson statistics carefully, and deonstrate the concept with simulated high-resolution X-ray spectral datasets of one-temperature, two-temperature and non-equilibrium plasma. Our proposed technique would assist scientists in finding important information that would otherwise be missed due to the unmanageable amount of data taken with future X-ray observatories.