📊 statistics

Identifying Gems from Roman RAPIDly

This paper introduces the RuBR machine learning model and a corresponding methodology for distinguishing genuine astronomical transients from spurious detections within the Roman Space Telescope's RAPID pipeline, demonstrating effective real-bogus classification strategies using simulated data to prepare for the mission's upcoming launch.

Original 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

Published 2026-06-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original 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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: A Cosmic "Needle in a Haystack" Problem

Imagine the Nancy Grace Roman Space Telescope as a super-powered camera about to be launched into space (planned for September 2026). Its job is to take millions of photos of the universe, looking for things that change or blink—like exploding stars (supernovae) or stars being eaten by black holes.

The problem? When you take a photo of the sky and compare it to an older photo to see what changed, your camera doesn't just show you the cool new stars. It also shows you "ghosts." These ghosts are caused by tiny dust specks on the lens, cosmic rays hitting the sensor, or just the computer getting confused. These are called "bogus" detections.

For every one real, exciting cosmic event, the computer might find 400 fake ones. It's like trying to find a single real diamond in a mountain of glass shards. If astronomers have to look at 400,000 alerts to find 1,000 real ones, they will never get any work done.

The Solution: A Smart Filter (RuBR)

The authors of this paper built a machine learning model called RuBR (Roman Real-Bogus). Think of RuBR as a highly trained security guard at the gate of a party. Its only job is to look at every alert the computer generates and decide: "Is this a real guest (a real star event), or is this an imposter (a glitch)?"

Since the telescope hasn't launched yet, the team couldn't train this guard on real data. Instead, they built a virtual universe (called OpenUniverse2024) and filled it with fake stars and fake glitches. They taught RuBR to spot the difference in this simulation.

How RuBR Works: The "Rotating" Detective

The paper highlights three special tricks RuBR uses to be so good:

The Rotating Glasses (Rotational Invariance):
Imagine looking at a star. It looks the same whether you hold your head straight, tilt it left, or turn it upside down. However, computers often get confused if an image is rotated.
- The Analogy: RuBR wears "rotating glasses." Before it makes a decision, it mentally spins the image 90 degrees, 180 degrees, and 270 degrees. It checks the object from all angles to make sure it's the same thing. This prevents the model from getting tricked by the direction of a glitch.
The Two-Brain System (Image + Features):
RuBR doesn't just look at the picture; it also reads a "report card" of numbers about the object.
- The Analogy: Imagine you are trying to identify a suspect.
  - Brain 1 (Image Encoder): Looks at the suspect's face (the picture of the star).
  - Brain 2 (Feature Encoder): Reads the suspect's ID card, which lists their height, weight, and shoe size (mathematical numbers like brightness and shape).
- RuBR combines both brains. If the face looks suspicious but the ID card says "this is a known glitch," RuBR catches it.
The "Filter" Strategy:
The telescope takes photos through different colored filters (like red, yellow, or infrared). A star looks different in red light than in blue light.
- The Analogy: RuBR learns to speak a different "dialect" for each color. It doesn't try to learn one rule for all colors; it learns specific rules for the red filter, specific rules for the blue filter, etc. This makes it much more accurate.

The Results: How Good is the Guard?

The team tested RuBR against other existing models (like Deep-HITS and Braai).

The Score: RuBR achieved a 90% precision rate. This means that out of every 100 alerts RuBR says are "Real," 90 of them actually are.
The Trade-off: It caught about 72% of the real events. It missed some faint ones, but it was very good at ignoring the fake ones.
Why it matters: Before this, the "Real-to-Bogus" ratio was terrible (1 real event for every 400 fakes). RuBR cleans up the list so astronomers can focus on the good stuff.

The Future: Adapting to the Real World

The paper admits a big challenge: The virtual universe they trained on isn't exactly like the real universe. When the telescope launches, the "ghosts" might look slightly different than the ones in the simulation.

To solve this, they tested a technique called Domain Adversarial Training.

The Analogy: Imagine you trained a dog to recognize a "Real Dog" using photos of Golden Retrievers. But when you take the dog to the park, it sees a Poodle. The dog might get confused.
The Fix: The authors taught the model to ignore the style of the photo (simulation vs. real) and focus only on the content (is it a glitch or a star?). They did this by playing a game where one part of the AI tries to guess "Is this a simulation?" and the other part tries to trick it by making the simulation look exactly like the real thing.
The Result: When they tested this on a different set of simulated data (acting as a stand-in for real data), the model that used this trick performed much better than the one that didn't.

Summary

This paper presents a smart, adaptable AI filter (RuBR) designed to clean up the massive stream of data from the upcoming Roman Space Telescope. By using a "two-brain" approach (pictures + numbers), spinning images to check for consistency, and learning to ignore the difference between simulations and reality, RuBR promises to help astronomers find real cosmic gems without getting buried in a mountain of digital trash.

Technical Summary: Identifying Gems from Roman RAPIDly (RuBR)

Problem Statement
The Nancy Grace Roman Space Telescope (Roman), scheduled for launch in September 2026, will conduct wide-field infrared surveys capable of detecting millions of astronomical transients. A critical bottleneck for such surveys is the "real-bogus" classification problem: image differencing pipelines (like RAPID) generate vast streams of candidate alerts, the majority of which are spurious detections (bogus) caused by subtraction artifacts, detector noise, or cosmic rays. While machine learning (ML) classifiers have successfully filtered these false positives in existing surveys, Roman presents a unique challenge: no real observational data exists yet to train these models. Consequently, there is a need for robust ML pipelines trained on simulated data that can effectively distinguish genuine transients from bogus detections and adapt to real-world data once the telescope is operational, without immediate reliance on ground-truth labels.

Methodology
The authors propose RuBR (Roman RAPIDly), a machine learning framework designed to classify transient candidates using a combination of image data and tabular features. The methodology encompasses three primary components:

Dataset Construction:
- The study utilizes the OpenUniverse2024 (OU24) simulations, which provide ~1.39 million simulated transient events across 70 deg², alongside locally injected transients on top of these images.
- The pipeline generates difference images (Science minus Reference) using algorithms like ZOGY or SFFT.
- Source detection is performed using DAOFind (chosen over SExtractor for higher completeness despite a higher false-positive rate).
- Training examples consist of $64 \times 64$ pixel cutouts of four image types (Science, Reference, Difference, and SCORR significance maps) centered on detections, augmented with five scalar features derived from DAOFind (magnitude difference, peak, sharpness, and two roundness values).
- The dataset is highly imbalanced, with a bogus-to-real ratio of approximately 400:1 for DAOFind detections.
Model Architecture:
- Image Encoder: A convolutional neural network (CNN) designed to be rotationally invariant. To ensure predictions remain consistent regardless of the source's orientation, the model averages embeddings over discrete rotations (multiples of 90°), preventing the model from latching onto spurious directional artifacts.
- Feature Encoder: A Multi-Layer Perceptron (MLP) processes the five scalar DAOFind features.
- Classifier: The image and feature embeddings are concatenated and passed through a fully connected layer with a sigmoid activation to output a probability score (the "rb-score") indicating the likelihood of a detection being real.
- Domain Adversarial Training (RuBRDA): To address the domain shift between simulated and real data, a domain predictor is added. Using a Gradient Reversal Layer (GRL), the model is trained to distinguish between simulated and real data domains while simultaneously forcing the feature extractor to learn domain-invariant representations. This allows the model to adapt to real data distributions without requiring ground-truth labels for the target domain.
Training Strategy:
- The models are trained using weighted binary cross-entropy to handle class imbalance.
- Filter-wise normalization is applied to account for distinct intensity distributions across different Roman filters.
- Three specific model variants are presented:
  - RuBRcomb: Trained and tested on a combined dataset of OU24 and local injections.
  - RuBRloc: Trained on local injections and tested on OU24.
  - RuBRDA: Trained on local injections with domain adaptation applied using a fraction of OU24 data as the target domain.

Key Results

Performance: On the combined test set, RuBRcomb achieved a Precision of 90.10% and a Recall of 71.68%, resulting in an F1-score of 0.7984. This outperformed several baseline architectures, including DenseNet121, VGG11, ResNet-18, Deep-HITS, and Braai.
Threshold Analysis: The model outputs a continuous confidence score. By adjusting the decision threshold, users can trade off between precision and recall. For instance, a threshold of 0.9 yields 98.34% precision but reduces recall to 58.47%.
Ablation Studies:
- Rotational Invariance: Removing this feature caused a significant drop in F1-score (from 0.7984 to 0.6218), confirming its necessity for robustness.
- Feature Encoder: Removing the scalar features reduced the F1-score to 0.7794, demonstrating that tabular features provide complementary information to image data.
- Normalization: Filter-wise normalization proved superior to standard global normalization.
Domain Adaptation: When tested on the OU24 dataset (simulating real data conditions) without ground-truth labels, the non-adapted model (RuBRloc) achieved only 36% precision. In contrast, the domain-adapted model (RuBRDA) improved precision to 49% and recall to 72%, demonstrating the efficacy of the adversarial training approach in bridging the simulation-to-real gap.

Significance and Claims
The paper claims that RuBR provides a viable pathway for the Roman mission to begin discovering reliable transients immediately upon launch, despite the absence of real training data. The primary significance lies in the domain-adversarial strategy, which allows the pipeline to adapt to the specific artifacts of real Roman observations without requiring immediate ground-truth labels. This is crucial for the early phases of the mission when human-verified labels are scarce.

The authors emphasize that while the current results are based on simulations, the methodology is designed to be robust against the high source density and crowding expected in the Galactic Bulge Time Domain Survey (GBTDS). They assert that the inclusion of confidence scores in the alert stream will empower astronomers to make informed decisions regarding follow-up observations. The work concludes that while the current recall is limited by subtraction artifacts near galaxy cores, the proposed framework offers a promising foundation for automated transient filtering in the Roman era, with future work focused on incorporating variable stars and refining the training set to exclude contaminated regions.