[Paper] Identifying Gems from Roman RAPIDly

Source: arXiv - 2606.05103v1

Overview

The paper introduces RuBR, a machine‑learning framework designed to separate real astronomical transients from spurious “bogus” detections in the upcoming Nancy Grace Roman Space Telescope data stream. Because Roman will generate millions of alerts before any real images exist, the authors devise a training strategy that works with simulated and injected sources, paving the way for an automated alert pipeline that can go live from day one.

Key Contributions

• RuBR model family – three variants (RuBR_comb, RuBR_loc, RuBR_DA) that handle different mixes of simulated (locally injected) and community‑generated (OpenUniverse2024) transients.
• Domain‑adaptation training – a novel approach that blends a small fraction of OpenUniverse data with locally injected sources to mitigate the “simulation‑real gap.”
• End‑to‑end pipeline integration – the models are built to slot directly into the RAPID transient‑detection pipeline, delivering real‑time real‑bogus scores.
• Comprehensive evaluation – extensive cross‑validation on both simulated and OpenUniverse test sets, showing > 95 % true‑positive rates at low false‑positive levels.
• Practical adaptation recipe – guidelines for fine‑tuning RuBR_comb on the first weeks of real Roman data when ground‑truth labels are unavailable.

Methodology

1. Data Generation

   • Locally injected transients: synthetic point‑source signals are added to raw Roman‑like images using the telescope’s expected PSF and noise characteristics.
   • OpenUniverse2024: a public benchmark dataset that mimics Roman’s cadence, cadence, and background variability.

2. Feature Extraction

   • Difference images (new – reference) are processed with a shallow CNN that learns spatial patterns (e.g., shape, flux distribution) indicative of real astrophysical events.
   • Additional handcrafted metrics (signal‑to‑noise, shape moments) are concatenated to the CNN embeddings, giving the model both learned and domain‑expert cues.

3. Model Variants

   • RuBR_comb: trained on a combined set of injected + OpenUniverse transients, aiming for the best overall performance.
   • RuBR_loc: trained only on injected data, then tested on OpenUniverse to quantify the simulation‑real gap.
   • RuBR_DA: uses domain‑adaptation – a small labeled slice of OpenUniverse is mixed with injected data, and a gradient‑reversal layer encourages the network to learn domain‑invariant features.

4. Training & Evaluation

   • Binary cross‑entropy loss with class‑weighting to counter the natural imbalance (many more bogus detections).
   • 5‑fold cross‑validation, ROC‑AUC, and precision‑recall curves are reported for each variant.

Results & Findings

Key takeaways:

• Adding even a modest amount of real‑like data (OpenUniverse) dramatically improves generalization.
• Domain‑adaptation recovers ~ 90 % of the performance loss seen in the pure‑simulation model.
• The combined model maintains low false‑positive rates, crucial for downstream follow‑up resources.

Practical Implications

• Ready‑to‑deploy alert filtering – Developers can plug RuBR_comb into Roman’s RAPID pipeline to automatically suppress bogus alerts, reducing the load on downstream human vetting and telescope scheduling systems.
• Transferable workflow – The same injection‑plus‑domain‑adaptation recipe can be applied to other upcoming surveys (e.g., Rubin LSST, Euclid) that suffer from a lack of early‑mission labeled data.
• Resource optimization – By cutting false alerts by an order of magnitude, observatories can allocate follow‑up time (spectroscopy, multi‑wavelength imaging) to truly novel transients, accelerating discovery of rare events like kilonovae or early supernova shock breakouts.
• Open‑source tooling – The authors release the data‑generation scripts and model checkpoints, allowing teams to fine‑tune the classifier on their own simulated pipelines or on early Roman commissioning data.

Limitations & Future Work

• Simulation fidelity – The injected transients assume perfect knowledge of the PSF and noise; any deviation in the actual Roman instrument could degrade performance.
• Label scarcity – While domain adaptation mitigates the gap, the approach still relies on a small set of real‑world labeled examples, which may be delayed during early operations.
• Model complexity – The current CNN is shallow to keep inference fast; exploring deeper architectures or transformer‑based vision models could further boost accuracy at the cost of latency.
• Extended variability – The study focuses on point‑source transients; future work should address extended sources (e.g., tidal disruption events) and periodic variables.

The authors plan to iterate on the image differencing front‑end, incorporate active‑learning loops for on‑the‑fly labeling, and test the pipeline on early commissioning data once Roman launches.

Authors

• Karan Gandhi
• Ashish A. Mahabal
• Jacob E. Jencson
• Russ R. Laher
• Ben Rusholme
• Lin Yan
• Ryan M. Lau
• Schuyler D. Van Dyk
• Mansi M. Kasliwal

Paper Information

• arXiv ID: 2606.05103v1
• Categories: cs.LG, astro-ph.IM, cs.CV, stat.ML
• Published: June 3, 2026
• PDF: Download PDF