Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Lei Hu, Tomás Cabrera, Antonella Palmese, Lifan Wang, Igor Andreoni, Xander J. Hall, Xingzhuo Chen, Jiawen Yang, Frank Valdes, Brendan O'Connor, Yuhan Chen

Abstract

We present a GPU-accelerated transient detection pipeline developed for time-domain surveys with the Dark Energy Camera (DECam). It enables real-time-capable image processing, incorporating science-driven candidate filtering to support rapid transient identification in time-critical observing programs. The pipeline serves as the core transient discovery engine for multiple long-term DECam programs, including the GW-MMADS gravitational-wave follow-up campaign and the DESIRT survey for intermediate-redshift transients with DESI synergy. The pipeline ingests calibrated imaging products from the DECam Community Pipeline and performs image differencing using the SFFT algorithm, coupled with CNN-based real-bogus classification, to produce science-ready transient alerts and light curves that are delivered to community brokers. We validate the pipeline using archival DECam data from the DESIRT survey. The real-bogus classifier achieves a completeness of $\sim$ 99\% of real transients while rejecting $\sim$ 96\% of subtraction artifacts, and the workflow typically reduces the candidate load to a manageable level for survey operations. With GPU acceleration, the typical processing time per DECam exposure is $\sim$ 50 s from calibrated image processing to alert generation using a modest allocation of computing resources.

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

Abstract

We present a GPU-accelerated transient detection pipeline developed for time-domain surveys with the Dark Energy Camera (DECam). It enables real-time-capable image processing, incorporating science-driven candidate filtering to support rapid transient identification in time-critical observing programs. The pipeline serves as the core transient discovery engine for multiple long-term DECam programs, including the GW-MMADS gravitational-wave follow-up campaign and the DESIRT survey for intermediate-redshift transients with DESI synergy. The pipeline ingests calibrated imaging products from the DECam Community Pipeline and performs image differencing using the SFFT algorithm, coupled with CNN-based real-bogus classification, to produce science-ready transient alerts and light curves that are delivered to community brokers. We validate the pipeline using archival DECam data from the DESIRT survey. The real-bogus classifier achieves a completeness of 99\% of real transients while rejecting 96\% of subtraction artifacts, and the workflow typically reduces the candidate load to a manageable level for survey operations. With GPU acceleration, the typical processing time per DECam exposure is 50 s from calibrated image processing to alert generation using a modest allocation of computing resources.
Paper Structure (19 sections, 1 figure, 1 table)

This paper contains 19 sections, 1 figure, 1 table.

Table of Contents

  1. Introduction
  2. Pipeline
  3. Stage 0: preparatory steps before observations
  4. Template image preparation
  5. Auxiliary catalog preparation
  6. Stage 1: Ingest new observations
  7. Stage 2: Image processing and transient detection
  8. Photometric calibration
  9. Image differencing
  10. Differential Photometry
  11. Real-bogus classification
  12. Alert generation
  13. Stage 3: Candidate filtering and science product generation
  14. Source association
  15. Candidate filtering
  16. ...and 4 more sections

Figures (1)

  • Figure 1: Schematic overview of the DECam transient detection pipeline. Colored boxes denote pipeline inputs, intermediate data products, processing components, and internal/external services. Arrows indicate data flow between modules, with solid arrows representing paths involving imaging data. The pipeline is divided into four stages, spanning pre-observation preparation (Stage 0), science data ingestion (Stage 1), image differencing and transient detection (Stage 2), and candidate filtering and science product generation (Stage 3).