The era of time-domain astronomy has brought new challenges and opportunities to X-ray observational astronomy. Recent discoveries indicate that fast extragalactic X-ray transients, planetary transits, and other exotic events can be extracted from rich X-ray telescope archives. Such anomalies have distinct signatures at the time series level, changes in patterns in the arrival times and energies of photons. Event files are inherently of variable length, posing an issue for data representation for deep learning applications. We compare two data representation methods for variable-length time series data: i) 2D histograms in the time and energy domain, and ii) connected graph representation. We train regression networks to a) predict posterior moments for global summaries of each time series and b) demonstrate an anomaly detection framework for unlabeled data. Trained networks are asked to predict archival summaries for a test data set, bypassing the need for an expensive processing pipeline. Outliers beyond a threshold in the predictive residual error distribution are flagged as potential anomalies for follow-up review. We test this method on a toy on-the-fly toy simulator and proceed to an application to real Chandra archival data. We present constructed Chandra graph and histogram data sets, as well as a toy model and regression network library written differentially in JAX, open-source tools for the X-ray community to enhance archival searches.
|