An interpretable and probabilistic approach to ACS COSI background modeling

Gamma-ray detectors aboard space missions operate in a background environment strongly influenced by spacecraft orientation and orbital position. The Anticoincidence System (ACS) of the Compton Spectrometer and Imager (COSI) satellite monitors the gamma-ray background and generates time series of detector counts with time bins of 50 ms. These measurements are used to identify transient events such as Gamma-Ray Bursts (GRBs) through onboard trigger algorithms. For this reason, accurate modeling of the background signal is essential in order to reliably distinguish transients from background fluctuations. This thesis proposes a probabilistic and interpretable machine learning framework for modeling the ACS background signal. Instead of predicting a deterministic estimate, the approach models the full conditional distribution of detector counts given the spacecraft state parameters. Several probabilistic regression models were investigated, including Skew-Normal regression and Gaussian Mixture Regression, trained via maximum likelihood to capture asymmetric and regime-dependent variability in the detector counts. To ensure interpretability, SHAP-based feature attribution methods were used to analyze the contribution of each input variable. The results indicate that background variations are primarily driven by spacecraft pointing directions and orbital geometry, consistent with physical effects such as Earth albedo and cosmic-ray induced background, and variations caused by spacecraft passage through the South Atlantic Anomaly (SAA). Kolmogorov–Arnold Networks were also explored as interpretable models capable of representing nonlinear relationships. Finally, the probabilistic model was integrated into a transient detection strategy using adaptive significance thresholds derived from the distributions. Experiments with simulated GRB signals demonstrate that the approach can identify significant excesses while accounting for background variability.