Title: Uncovering True Signals from Roman’s RAPID Data

Abstract: With a planned launch no later than September 2026, the Nancy Grace Roman Space Telescope (Roman) is poised to execute wide-field infrared imaging surveys characterized by exceptional spatial resolution and temporal cadence. These capabilities are expected to yield the detection of millions of astronomical transients. To ensure the telescope can immediately identify reliable transients and variable objects upon becoming operational, automated alert-generation pipelines must be established beforehand. However, the absence of actual Roman data has historically complicated the development of these systems. This study introduces $RuBR$, a machine learning model alongside a broader methodology designed to differentiate authentic transient and variable detections from false positives (bogus events) within the RAPID pipeline. Specifically, we detail three iterations of this approach: $RuBR_{comb}$, which is trained and evaluated on a merged dataset of locally injected and OpenUniverse2024 transients; $RuBR_{loc}$, trained exclusively on locally injected transients but tested against OpenUniverse2024 data; and $RuBR_{DA}$, which integrates locally injected transients with a subset of OpenUniverse2024 transients using a domain-adaptation training regime. This framework establishes a foundation for adapting the $RuBR_{comb}$ model to real observational data, addressing the challenge of missing ground-truth labels during the initial stages of the Roman mission. Although improvements to the image differencing pipeline are ongoing, our experimental findings confirm the efficacy of the proposed method and its potential for achieving robust real-bogus classification in the upcoming Roman era.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC