Autoencoding Poisson Processes for Unsupervised X-ray Sources Learning

Recent advancements in massive serendipitous X-ray event detections necessitate efficient automatic tools to study the properties of X-ray sources. However, the variable-length time series in X-ray eventfiles as well as the inherent randomness of photon arrivals pose challenges to effective approaches that characterize relevant time-domain and spectral anomalies in an automatic fashion, directly from the event files. Although different machine learning models have been proposed to classify and represent X-ray sources, they tend to oversimplify the problem by either using fixed-length inputs of hand-designing features-therefore losing input information-or using a reconstruction based loss that ignores the stochastic nature of arrivals. We propose a carefully designed Poisson Process AutoEncoder (PPAE) that addresses both challenges. PPAE consists of an auto-regressive encoder that maps the variable-length input eventfiles to latent features and a neural field decoder that translates those latent features to continuous Poisson rate functions. We train PPAE with log likelihoods of Poisson arrivals as the loss function to precisely capitalize the statistical nature of these events. Moreover, PPAE learns to represent eventfiles in an unsupervised manner in a low dimension, thereby offering great flexibility in various downstream tasks such as classification and anomaly detection. We apply PPAE on ~100,000 sources in the Chandra archive to demonstrate its effectiveness. We will release the code and learned embeddings for further advancing the ML-driven discovery in the astronomy community.