Physics-inspired spatiotemporal-graph AI ensemble for the detection of higher order wave mode signals of spinning binary black hole mergers

TL;DR

This work addresses the challenge of detecting gravitational-wave signals with higher-order modes from quasi-circular, spinning binary black holes using a physics-informed AI framework. It introduces a spatiotemporal-graph AI ensemble that fuses a hybrid dilated convolution network for temporal structure with a graph neural network for robust multi-detector fusion, applied to higher-order modes such as $(l,|m|)= {(2,2),(2,1),(3,3),(3,2),(4,4)}$ and mode mixing. Trained on 1.2 million waveforms with synthetic noise and validated on large-scale HPC, the approach achieves state-of-the-art detection performance, scales to hundreds of GPUs, and processes years of data rapidly; on O3b data it identifies 6 real HLV events with zero false positives. Compared with traditional pipelines, the method significantly reduces false alarms while maintaining sensitivity, demonstrating a scalable, physics-informed path toward real-time GW discovery across increasingly complex waveform manifolds.

Abstract

We present a new class of AI models for the detection of quasi-circular, spinning, non-precessing binary black hole mergers whose waveforms include the higher order gravitational wave modes $(l, |m|)=\{(2, 2), (2, 1), (3, 3), (3, 2), (4, 4)\}$, and mode mixing effects in the $l = 3, |m| = 2$ harmonics. These AI models combine hybrid dilated convolution neural networks to accurately model both short- and long-range temporal sequential information of gravitational waves; and graph neural networks to capture spatial correlations among gravitational wave observatories to consistently describe and identify the presence of a signal in a three detector network encompassing the Advanced LIGO and Virgo detectors. We first trained these spatiotemporal-graph AI models using synthetic noise, using 1.2 million modeled waveforms to densely sample this signal manifold, within 1.7 hours using 256 A100 GPUs in the Polaris supercomputer at the ALCF. Our distributed training approach had optimal performance, and strong scaling up to 512 A100 GPUs. With these AI ensembles we processed data from a three detector network, and found that an ensemble of 4 AI models achieves state-of-the-art performance for signal detection, and reports two misclassifications for every decade of searched data. We distributed AI inference over 128 GPUs in the Polaris supercomputer and 128 nodes in the Theta supercomputer, and completed the processing of a decade of gravitational wave data from a three detector network within 3.5 hours. Finally, we fine-tuned these AI ensembles to process the entire month of February 2020, which is part of the O3b LIGO/Virgo observation run, and found 6 gravitational waves, concurrently identified in Advanced LIGO and Advanced Virgo data, and zero false positives. This analysis was completed in one hour using one A100 GPU.

Figures (6)

• Figure 1: Classification performance of AI ensembles to detect quasi-circular, spinning, non-precessing black hole mergers. Classification performance in terms of the Receiver Operating Characteristic (ROC) curve, and the Precision-Recall (PR) curve for ensembles that include 2 AI models (top row), 3 AI models (center row), and 4 AI models (bottom row). These results were produced using a decade-long gravitational wave test set, for the Advanced LIGO and Advanced Virgo three detector network, in which we injected 300,000 modeled binary black hole waveforms that cover a broad SNR range.
• Figure 2: Classification performance of AI ensemble throughout February 2020. Our 6 AI model ensemble was able to identify 6 HLV events in February 2020 with no false positives. The detected events from left to right are: GW200129_065458_1264314069, GW200208_130117_1265200048, GW200209_085452_1265271663, GW200216_220804_1265924055, GW200219_094415_1266138626, and GW200224_222234_1266616125.
• Figure 3: Architecture of hybrid dilated convolution network (HDCN). Each detector input will be processed by an individual HDCN, which consists of the N blocks of pre-processing time-distributed convolution layer, dilated convolution layer, and the post-processing time-distributed convolution layer. N depends on the receptive field desired. Here we use N=11 to cover a half second receptive field. Each block has residual structure and skip connection to the output embedding.
• Figure 4: Spatiotemporal-graph AI model structure. (a) Graph neural network (GNN) structure, (b) 3 prediction embeddings given by the hybrid dilated convolution network (HDCN) block, (c) Geometric visualization for prediction embeddings of the three time series produced by the Advanced LIGO and Advanced Virgo detectors.
• Figure 5: Distributed training in the Polaris supercomputer. We developed novel methods to complete the training of spatiotemporal-graph AI models within 1.7 hours using 256 A100 NVIDIA GPUs. Our optimization method exhibits strong scaling up to 512 NVIDIA A100 GPUs.