Researchers have introduced CASPER, a lightweight machine learning pipeline that combines a residual neural network classifier with an interpretability layer to detect gravitational wave signals, and explain why it made each decision. The work tackles a growing challenge in gravitational wave astronomy: as detection rates surge past 390 confirmed events (as 1ban.news covered in the GWTC-5.0 catalog release earlier this month), the need for fast, reliable, and interpretable analysis tools has become urgent.
The paper, submitted to arXiv on June 15, was authored by R. Rai, R. Verma, and Somya from institutions in India.
### Beyond Matched Filtering
Traditional gravitational wave detection relies on matched filtering, a technique that cross-correlates detector data against a library of pre-computed waveform templates. While effective and well-proven across nearly a decade of LIGO-Virgo-KAGRA (LVK) operations, matched filtering has fundamental limitations. It requires extensive template banks, is computationally expensive for long-duration signals, and critically, provides no insight into which specific features of the signal drove the detection decision.
Deep learning alternatives have shown competitive sensitivity, but many have struggled with real-world deployment. The authors identify four persistent failure modes: class overlap (where noise and signal features are difficult to distinguish), imbalanced class weighting (rare signals among vast noise), limited sample variation in training, and train-test mismatch that causes poor generalization when models trained on synthetic data encounter real detector noise.
### CASPER: ResNet Meets Shapley Values
The CASPER pipeline (Classification with Attribution via ShaPlEy in Residual neural networks) addresses these issues with an end-to-end approach. A residual convolutional neural network (ResNet) serves as the classifier, trained on 260 distinct events from the Gravitational Wave Open Science Centre (GWOSC) spanning a signal-to-noise ratio (SNR) range of 7 to 42. Critically, the model was trained on data from both the LIGO Hanford (H1) and LIGO Livingston (L1) detectors using only real events; no synthetic augmentation was employed, directly addressing the train-test mismatch problem.
Focal Loss was used to handle class imbalance by down-weighting easy examples and focusing training on hard-to-classify cases. Platt Calibration was applied to improve probability estimates and generalization performance.
The classifier achieves an Area Under the Curve (AUC) of 91 percent with a low false alarm rate. But the key innovation lies in the FastSHAP explainer, which generates attribution maps that reconstruct the complete chirp morphology of detected signals: the characteristic rising-frequency sweep that marks a gravitational wave event. These maps provide a visual interpretation of which signal features drove the model’s classification decision, something matched filtering cannot offer.
### A Lightweight Alternative
The complete CASPER pipeline contains fewer parameters than standard deep learning models and runs on a standard central processing unit (CPU) and requires no graphics processing unit (GPU) or specialized hardware. This makes it potentially suitable for real-time deployment at observatory sites where computing resources may be limited.
“FastSHAP attribution maps recover the complete chirp morphology and provide detailed maps for a visual interpretation of the decision,” the authors write.
### Context: Gravitational Wave Astronomy at Scale
The CASPER paper arrives at a moment when gravitational wave astronomy is transforming from a discovery science into a statistical enterprise. The latest LVK catalog, GWTC-5.0, released on May 26, added 161 new black hole merger signals detected during the second half of the fourth observing run (O4b, April 2024 through January 2025), bringing the total confirmed events to 390 since the first detection in 2015. As 1ban.news reported in our coverage of the GWTC-5.0 release, the fourth observing run alone now accounts for roughly 75 percent of all gravitational wave events ever detected.
With the next observing run (O5) expected to push detection rates to thousands of events per year, aided by detector sensitivity upgrades and the eventual addition of the LIGO India observatory, the need for lightweight, interpretable, and generalizable analysis tools will only grow. CASPER represents a step toward an AI analysis pipeline that can handle the coming data flood while still telling scientists why it reached each conclusion.