A GPU-Accelerated Transient Detection Pipeline for DECam Time-Domain Surveys

This paper presents a GPU-accelerated transient detection pipeline for DECam time-domain surveys that integrates image differencing and CNN-based classification to achieve real-time processing, high completeness in identifying real transients, and efficient artifact rejection for rapid scientific discovery.

The Gist

Imagine the night sky as a giant, bustling city that never sleeps. Astronomers use powerful cameras, like the Dark Energy Camera (DECam), to take snapshots of this city every few days, hoping to catch a glimpse of something new: a star exploding, a black hole eating a star, or a distant galaxy flashing.

But here's the problem: taking these photos is like trying to find a single new firefly in a stadium full of blinking lights, dust motes, and camera glitches. Most of the "new" things the computer sees aren't real; they are just shadows, dust on the lens, or cosmic rays hitting the sensor. These are called "bogus" detections. Real discoveries are called "real" transients.

This paper introduces a super-fast, smart system designed to act as a high-tech security guard for the night sky. It's a pipeline that automatically sorts through thousands of images, finds the real fireworks, and ignores the dust, all in about 50 seconds.

Here is how it works, broken down into simple steps:

1. The "Before the Game" Prep (Stage 0)

Before the camera even takes a picture, the system gets ready. Think of this like a detective preparing a "Wanted" poster and a map of the neighborhood.

• The Template: The system grabs an old, high-quality photo of the same patch of sky (taken weeks or months ago) to use as a reference. This is the "perfect" version of the sky without any new events.
• The Cheat Sheet: It loads a database of known stars, galaxies, and even asteroids so it knows what should be there.

2. The "New Photo" Arrival (Stage 1)

When the camera takes a new picture, the system instantly downloads it. It's like a mailman dropping off a fresh newspaper. The system organizes this new photo next to the old "template" photo, lining them up perfectly, pixel by pixel.

3. The "Subtraction Magic" (Stage 2)

This is the most critical part. The system takes the New Photo and subtracts the Old Photo.

• The Analogy: Imagine taking a photo of a living room, then taking another photo 10 minutes later. If you subtract the first photo from the second, everything that stayed the same (the sofa, the lamp, the rug) disappears. You are left with only what changed: a cat walking across the floor or a new toy on the table.
• The Tech: In astronomy, this "subtraction" is done using a super-fast math trick called SFFT that runs on GPUs (the same powerful chips used in video game computers). This allows the system to do the math incredibly fast, spotting even the faintest new "cat" (transient) in the sky.

4. The "Smart Filter" (The AI Detective)

After subtraction, the system has a list of thousands of "new" spots. But many are still fake (dust, bad pixels, or math errors).

• The CNN Classifier: The system uses a Convolutional Neural Network (CNN), which is a type of Artificial Intelligence trained to look at the "stamps" (tiny crop-out images) of these new spots.
• Training: The AI was trained on millions of examples, learning to tell the difference between a real exploding star and a cosmic ray glitch. It's like a seasoned detective who can spot a fake ID in a split second.
• The Score: It gives every candidate a "Realness Score" from 0 to 1. If the score is high, it's likely real. If it's low, it's likely garbage.

5. The "Final Report" (Stage 3)

The system filters out the junk and keeps only the most promising candidates.

• The Light Curve: For the real ones, it builds a "story" of how bright they are over time (a light curve).
• The Alert: It packages this information into a neat report and sends it to astronomers around the world (via systems like TNS and SkyPortal) so they can point their telescopes at the new object immediately.

Why is this paper special?

• Speed: In the past, processing a single image might take minutes or hours. This system does it in about 50 seconds. This is crucial for catching things that happen fast, like the flash of a gravitational wave event.
• Accuracy: It catches 99% of the real events while throwing away 96% of the fake ones.
• Scalability: It's built to handle the massive amount of data coming from modern surveys, acting as the engine that powers several major scientific projects.

In a nutshell: This paper describes a super-fast, AI-powered "sky vacuum cleaner" that sucks up the noise and dust, leaving only the beautiful, real cosmic fireworks for astronomers to study. It turns a mountain of data into a manageable list of discoveries, allowing humanity to react to the universe in near real-time.

Technical Summary

Here is a detailed technical summary of the paper "A GPU-Accelerated Transient Detection Pipeline for DECam Time-Domain Surveys."

1. Problem Statement

The Dark Energy Camera (DECam), mounted on the 4-m Blanco Telescope, supports diverse time-domain astronomy programs, including gravitational-wave (GW) follow-up (GW-MMADS) and intermediate-redshift transient surveys (DESIRT). These programs require:

• Real-time or near-real-time processing: Critical for time-critical events like GW counterparts (kilonovae) where rapid identification is essential.
• High-volume data handling: DECam produces large datasets (62 CCDs, ~3 deg² field of view) requiring efficient subtraction and candidate filtering.
• Robust artifact rejection: Difference imaging often produces spurious detections (subtraction artifacts, cosmic rays, PSF mismatches) that overwhelm human vetting efforts.
• Flexibility: The pipeline must handle both scheduled survey cadences and event-driven Target-of-Opportunity (ToO) observations with minimal tuning.

Existing pipelines often struggle to balance computational speed with the sensitivity required to detect faint transients while maintaining high purity.

2. Methodology

The authors developed a four-stage, GPU-accelerated pipeline designed to process calibrated DECam imaging products from the DECam Community Pipeline (CP).

Pipeline Architecture

• Stage 0: Pre-observation Preparation:
  • Initializes based on an observing plan (coordinates, filter, date).
  • Constructs a local metadata mirror of the NOIRLab archive for rapid template selection.
  • Prepares auxiliary catalogs (Legacy Survey, Pan-STARRS1, Gaia, NED-LVS) for flux calibration, star/galaxy separation, and host-galaxy redshift assignment.
  • Template Generation: Supports two modes:
    • Single-exposure: Selects the deepest archival exposure matching the filter and pointing.
    • Multi-exposure: Constructs synthetic templates by resampling and aligning multiple archival exposures to a common WCS when no single deep image exists.
• Stage 1: Science Data Ingestion:
  • Automatically downloads calibrated science images (.ooi) and quality masks (.ood) from the NOIRLab archive.
  • Organizes data hierarchically by program, pointing, and exposure.
• Stage 2: Image Differencing & Transient Detection (Core):
  • Alignment: Reprojects science images to the template WCS using SWarp.
  • Image Subtraction: Utilizes the GPU-accelerated Saccadic Fast Fourier Transform (SFFT) algorithm. This method models the convolution kernel in the Fourier domain, accounting for spatial PSF variations and sub-pixel misalignments without requiring extensive user parameters.
  • Photometry: Performs differential photometry on difference images using SExtractor.
  • Real-Bogus Classification: Applies a rotation-invariant Convolutional Neural Network (CNN) to classify candidates.
    • Input: Three-channel image stamps (Science, Template, Difference).
    • Architecture: Hybrid model based on ResNet-18 with rotation-invariant layers.
    • Training: Trained on DESIRT data augmented with simulated transient injections (137k real, 157k artifacts).
  • Alert Generation: Filters candidates based on a configurable CNN score threshold (default "hot" threshold = 0.3 for sensitivity; "cold" threshold = 0.6 for purity).
• Stage 3: Candidate Filtering & Product Generation:
  • Object Association: Links alerts across exposures into unique objects based on sky position (1.9" matching radius).
  • Filtering: Removes stellar variability (using Gaia proper motion), known minor planets (MPC database), and enforces multi-detection requirements.
  • Output: Generates finding charts (triplet stamps), light curves (forced aperture photometry), and packages alerts into ZTF-style Avro format for brokers (e.g., BOOM, SkyPortal) and the Transient Name Server (TNS).

3. Key Contributions

• GPU-Accelerated SFFT: Implementation of the SFFT algorithm on GPUs (NVIDIA Tesla A100) enables high-throughput image subtraction, a significant bottleneck in traditional CPU-based pipelines.
• Hybrid Template Strategy: A flexible approach supporting both single-deep and multi-exposure synthetic templates, ensuring sensitivity across diverse survey strategies (e.g., ToO vs. regular survey).
• Advanced CNN Classifier: A rotation-invariant CNN specifically trained on DECam data with simulated injections, achieving high completeness and purity in distinguishing real transients from subtraction artifacts.
• End-to-End Automation: A fully automated workflow from raw plan ingestion to science-ready light curves and broker alerts, minimizing human intervention.
• Scalability: Designed to handle the specific cadence and volume of DECam time-domain surveys, capable of processing an exposure within the interval between consecutive observations.

4. Results

The pipeline was validated using archival data from the DESIRT survey (2025A semester, ~3,642 exposures).

• Artifact Rejection Performance:
  • Default Configuration (Hot Threshold, Score > 0.3): Retains ~98.9% of real transients while rejecting ~95.8% of bogus detections.
  • High-Purity Configuration (Cold Threshold, Score > 0.6): Retains ~97.5% of real transients while rejecting ~98.9% of bogus detections.
• Candidate Reduction:
  • Starting with ~910,000 alerts from difference images, the pipeline reduced the sample to 2,737 objects (0.3%) after automated cuts.
  • Human vetting identified 76 promising candidates from this reduced set, demonstrating effective noise suppression.
• Computational Efficiency:
  • Stage 2 Latency: The core image differencing and detection stage processes a single DECam exposure in approximately 50 seconds on a cluster with 64 CPU cores and 2 NVIDIA Tesla A100 GPUs.
  • This meets the requirement of processing within the ~1-minute interval between consecutive DECam exposures, enabling near-real-time operations.
• Case Study: The pipeline successfully identified and characterized AT 2025ifl, a Type Ia supernova at $z=0.127$, generating high-quality light curves and finding charts.

5. Significance

This work provides a critical infrastructure upgrade for DECam time-domain astronomy. By leveraging GPU acceleration and deep learning, the pipeline solves the latency and scalability challenges inherent in processing wide-field survey data.

• Enabling Multi-Messenger Astronomy: The speed and sensitivity make it ideal for GW follow-up campaigns (GW-MMADS), where rapid identification of faint kilonovae is crucial.
• Synergy with DESI: The pipeline supports the DESIRT survey's goal of characterizing intermediate-redshift transients in coordination with DESI spectroscopic data.
• Community Impact: By delivering standardized, broker-compatible alerts (Avro) and light curves, the pipeline facilitates rapid community follow-up and maximizes the scientific return of DECam observing time.
• Future-Proofing: The modular, GPU-native design serves as a blueprint for next-generation time-domain surveys (e.g., LSST/Rubin) where data volumes will be even larger.