Argonne National Laboratory accelerates cosmic discovery with AI-powered RADAR framework

Overview

Argonne National Laboratory and university collaborators developed RADAR, a federated, privacy-enhancing framework that lets observatories coordinate gravitational-wave and radio follow-up without moving or exposing proprietary data. Running site-local AI inference on NVIDIA GPUs across the Polaris, Delta and DeltaAI supercomputers, RADAR processed over an hour (4,096 seconds) of Advanced LIGO data in under 4.5 minutes, a 5-10x speedup versus CPU workloads. The RADAR paper has been accepted for publication in The Astrophysical Journal Supplement Series.

The challenge

In multi-messenger astrophysics, rapid, locally executed data processing is essential for identifying gravitational-wave events, but heterogeneous data-sharing policies across the radio astronomy community — with proprietary periods ranging from immediate public dissemination to embargoes lasting until peer-reviewed publication — made centralized data movement impractical, historically limiting broad participation in radio follow-up campaigns and leaving valuable measurements siloed.

The solution

Argonne National Laboratory and collaborating institutions developed RADAR, a federated, privacy-enhancing framework that identifies gravitational-wave events using site-local AI inference and matches detections with public LIGO-Virgo-KAGRA 'superevent' alerts, publishing a trigger to the cloud-hosted Octopus messaging fabric when a match is found. RADAR uses an AI-powered parser to convert unstructured GCN circulars into structured metadata, then runs Dingo-BNS for gravitational-wave parameter estimation and afterglowpy for federated radio afterglow modeling, comparing and combining posteriors to refine source parameters while each observatory retains ownership of its local datasets. The framework runs on NVIDIA A100, A40 and GH200 GPUs across the Polaris, Delta and DeltaAI supercomputers.

Reported business value

Integrating NVIDIA GPUs on the Polaris, Delta and DeltaAI supercomputers, RADAR's AI module for signal detection processed over an hour (4,096 seconds) of Advanced LIGO data in under 4.5 minutes, a 5-10x speedup versus CPU workloads. The RADAR paper has been accepted for publication in The Astrophysical Journal Supplement Series.