🔭 astrophysics

The SPOTLIGHT Multibeam Real-Time Transient Detection System

This paper presents the design, implementation, and initial performance of SPOTLIGHT, a GPU-accelerated real-time transient detection system deployed on the upgraded Giant Metrewave Radio Telescope (uGMRT) that successfully detected 2,870 bursts from 42 known sources while processing up to 2,000 beams to enable wide-field, low-frequency Fast Radio Burst discovery.

Original authors: Ujjwal Panda, Jayanta Roy, Kshitij Bane, Chahat Dudeja, Sridhar Gajendran, Param Joshi, Sanjay Kudale, Arpan Pal, Raghav Wani, Karel Adamek, Wes Armour, Kenil Ajudiya, Jayaram Chengalur, Jyotirmoy Das

Published 2026-06-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ujjwal Panda, Jayanta Roy, Kshitij Bane, Chahat Dudeja, Sridhar Gajendran, Param Joshi, Sanjay Kudale, Arpan Pal, Raghav Wani, Karel Adamek, Wes Armour, Kenil Ajudiya, Jayaram Chengalur, Jyotirmoy Das, Nishant Pradeep Deo, Deepak Bhong, Shelton Gnanaraj, Santaji N. Katore, Mekhala Muley, Harshavardhan Reddy

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: Catching Cosmic "Fireflies"

Imagine the universe is a dark forest at night. Most of the time, it's quiet. But occasionally, a tiny, incredibly bright firefly flashes for just a split second and then vanishes. In astronomy, these "fireflies" are called Fast Radio Bursts (FRBs). They are mysterious, short-lived bursts of radio energy coming from deep space.

The problem is that these flashes are so fast and faint that catching them is like trying to spot a specific firefly in a storm while wearing blinders. You need a super-sensitive net that can sweep a huge area of the sky instantly.

The Solution: The SPOTLIGHT System

The SPOTLIGHT project is a new high-tech "net" built to catch these cosmic fireflies. It is attached to the Giant Metrewave Radio Telescope (uGMRT) in India, which is essentially a giant ear listening to the universe.

Instead of building a new telescope, SPOTLIGHT is a massive computer system (a "backend") that works alongside the telescope's normal observations. Think of it like a dedicated security guard who watches the security camera feeds while the main security team is busy doing other tasks. This is called a "commensal" survey—it does its job without slowing down the telescope's regular work.

The Hardware: A Super-Computer Brain

To process the massive amount of data coming from the telescope, SPOTLIGHT uses a dedicated supercomputer cluster.

The Muscle: It has 90 NVIDIA A100 GPUs. If you imagine a standard computer as a bicycle, these GPUs are like a fleet of Formula 1 race cars working in perfect sync.
The Scale: The system is so powerful it can process data for 2,000 different "beams" (directions in the sky) all at the same time. It's like having 2,000 pairs of eyes scanning the sky simultaneously.

How It Works: The Assembly Line

The paper describes the system as a two-stage assembly line that filters out noise to find the real signals.

1. The "Brute Force" Search (aamulti)
Imagine you are looking for a specific word in a library of millions of books, but the pages are scrambled by static noise.

Dedispersion: Radio signals from space get smeared out by space dust (plasma) as they travel. The system has to "un-smeared" the signal, like a photo editor fixing a blurry picture. It tries thousands of different "un-smeared" settings to see which one makes the signal clear.
The Search: Once the signal is clear, the system looks for "single pulses" (the flashes). It uses a "matched filter," which is like using a template to stamp out the exact shape of a flash. If the shape matches, it flags it.

2. The "Smart Filter" (spltpipe)
The first stage finds millions of potential flashes, but most of them are just interference (like a microwave oven or a car radio). This second stage is the "smart filter" that cleans up the list.

Clustering: It groups similar flashes together. If a flash appears in one direction but not its neighbors, it might be real. If it appears everywhere at once, it's likely local interference (RFI).
The "Anti-Coincidence" Trick: The system splits the sky into different groups. If a "flash" shows up in two groups that are looking at completely different parts of the sky, the system knows it's a fake signal (interference) and throws it away.
AI Classification: The remaining candidates are fed into an AI (Machine Learning) trained to recognize the specific "fingerprint" of a real Fast Radio Burst. It looks for a specific "bow-tie" pattern in the data that only real cosmic signals have.

The Safety Net: Testing with Fake Fireflies

How do they know the system works? They built a tool called arachne.

The Analogy: Imagine a magician who secretly plants a fake firefly in the forest to see if the security guard spots it.
The Reality: arachne injects synthetic, fake radio bursts directly into the live data stream. If the system catches the fake burst, the scientists know the system is working perfectly.

The Results: What Did They Find?

During its first few months of operation (late 2025 to mid-2026), the system was tested and then fully deployed.

The Catch: It successfully detected 2,870 bursts from 42 known sources (mostly repeating radio sources we already knew about).
Sensitivity: It is sensitive enough to catch very faint flashes (about 0.2 Jy ms), which matches the scientists' predictions.
Speed: The system is fast enough to process the data in "real-time." This means that the moment a burst happens, the system knows about it immediately, rather than waiting days to analyze the data later.

Why This Matters

The paper concludes that SPOTLIGHT is a major step forward. It proves that we can build a system that is:

Fast: It catches events as they happen.
Wide: It looks at a huge patch of sky at once.
Smart: It uses AI and massive computing power to ignore noise and find the real signals.

This setup provides the foundation for future surveys that will help us understand where these mysterious cosmic fireflies come from and what they tell us about the universe.

Technical Summary: The SPOTLIGHT Multibeam Real-Time Transient Detection System

Problem Statement
Fast Radio Bursts (FRBs) are enigmatic, millisecond-duration radio transients of extragalactic origin. While their discovery rate has increased significantly, particularly with wide-field instruments like CHIME/FRB, the majority of detections lack precise localization (precision $\lesssim 5$ arcseconds) required to associate them with host galaxies. This limitation hinders the study of FRB progenitors, local environments, and their utility as cosmological probes. Existing interferometric surveys often face a trade-off between field-of-view (FoV) and sensitivity, or lack the real-time processing capabilities necessary to capture high-time-resolution data for transient localization. There is a critical need for a system that combines the high sensitivity and localization capabilities of an interferometer with a wide FoV and real-time processing to conduct commensal searches for FRBs.

Methodology and System Architecture
The SPOTLIGHT (Survey for sPoradic radiO bursTs via a commensaL multI-beam GPU-powered HPC at the GMRT) project addresses this by leveraging the upgraded Giant Metrewave Radio Telescope (uGMRT). The system is built on a dedicated High-Performance Computing (HPC) facility, "Param Brahmand," comprising 60 Rudra servers (indigenously developed by C-DAC) equipped with 90 NVIDIA A100 GPUs and 2 PB of storage.

The core of the system is a real-time transient search pipeline designed to process up to 2000 post-correlation (PC) beams, currently operating with 640 beams at a time resolution of 1.3 ms and 4096 frequency channels. The pipeline architecture is divided into two asynchronous, GPU-accelerated modules to maximize hardware utilization:

aamulti (AstroAccelerate-powered): This module performs the initial brute-force dedispersion and single-pulse search. It utilizes the AstroAccelerate library to handle dispersion measures (DM) up to 2000 pc cm $^{-3}$ across uGMRT bands (300–1460 MHz). It employs an overlap-read strategy to recover signals crossing block boundaries. The search uses matched filtering with boxcar templates to identify single pulses, followed by peak filtering using 3 $\times$ 3 and 11 $\times$ 11 stencils to reduce candidate counts.
spltpipe (Candidate Optimization): This module refines the candidates from aamulti through a multi-stage filtering process:
- Clustering: Uses DBSCAN to group candidates in the DM-time plane, selecting the highest Signal-to-Noise (S/N) peak per cluster.
- Thresholding: Applies cuts on S/N (>10), DM (>10 pc cm $^{-3}$ ), and width (<100 ms).
- Coincidence/Anti-coincidence Filtering: Distributes beams across nodes to distinguish between astrophysical signals (present in specific beams) and Radio Frequency Interference (RFI) (present in all or many beams simultaneously).
- Classification: Extracts features (dedispersed dynamic spectrum and DM transform) using the candies package, which offers improved feature generation over the previous your package. These features are classified using FETCH, a Convolutional Neural Network (CNN) trained on pulsar and FRB data.
- Triggering: Candidates passing a 50% probability threshold trigger the dumping of high-resolution visibility (1.3 ms) and baseband data for offline imaging and analysis.

The system also includes an incoherent array (IA) beam search (aa.single and splt.single) for wider FoV coverage and a real-time injection framework, arachne, which injects synthetic bursts into the data stream to validate pipeline performance continuously.

Key Contributions

Scalable Real-Time Pipeline: The development of a GPU-accelerated pipeline capable of processing 640 beams in real-time (or better) with a DM search range up to 2000 pc cm $^{-3}$ .
Advanced Feature Extraction and Classification: The introduction of candies, a GPU-optimized package for generating dynamic spectrum and DM transform features, which demonstrates a 1 $\times$ to 3 $\times$ speedup over your and significantly improves the accuracy of the FETCH classifier, particularly for low S/N events.
Real-Time Validation Framework: The implementation of arachne for injecting realistic, quantized synthetic bursts directly into the live data stream, ensuring continuous performance monitoring without interrupting science observations.
Commensal Operation: Successful integration of the backend with routine uGMRT observations, allowing for simultaneous high-sensitivity transient searches without sacrificing primary science goals.

Results
During its initial deployment in uGMRT Cycle 49 and Cycle 50 (October 2025 – June 2026), the SPOTLIGHT pipeline processed data from 640 beams. The system detected 2,870 bursts from 42 known astrophysical sources.

Sensitivity: The detected bursts span a wide range of dispersion measures and fluences, reaching the predicted survey threshold of $\sim 0.2$ Jy ms (for an S/N of 10).
Performance: The pipeline achieved near real-time processing factors for Band 3 (300–500 MHz) and Band 4 (550–750 MHz). Band 5 (1060–1460 MHz) showed slightly slower-than-real-time performance, attributed to lower dispersive delays making RFI discrimination more challenging.
Sky Coverage: The system covered approximately 277.77 square degrees across all bands, with the majority of coverage in Band 3.

Significance
The paper claims that SPOTLIGHT establishes a scalable framework for wide-field, low-frequency transient discovery and localization. By successfully combining high sensitivity, large bandwidth, and real-time processing on a dedicated HPC cluster, the system provides a technological foundation for next-generation radio transient surveys. The detection of thousands of bursts from known sources validates the pipeline's sensitivity and robustness, demonstrating its capability to operate commensally with routine telescope operations. The system represents one of the largest GPU-powered commensal transient-search systems currently operating at low radio frequencies.