Original title: A Fast and Scalable Transformer Pipeline for Binary Black Hole Detection
Chayan Chatterjee, Abigail Petulante, Haowei Fu, Yang Hu +2 · August 2026
Scientists expect to spot many more crashing black holes soon, but current search methods need fast and scalable alternatives. In this paper, researchers introduce CASTOR, a system based on transformers—the same kind of computer model used for text and speech—to find gravitational waves from colliding binary black holes in Advanced LIGO data. CASTOR searches data from multiple detectors at once. It can calculate its false-alarm rate using time slides cheaply, without needing to rerun the neural network over and over. When tested, CASTOR ranks among the most sensitive machine-learning pipelines on challenge data. It spots most confident events from the GWTC-3 catalog that fall within its training range. It also beats GW-Whisper, an adapted audio model, in sensitivity while cutting the computing cost of background estimation by a factor of 20. This makes it a practical tool for future searches.
You are a lifeguard scanning a crowded beach for someone in trouble. Normally you compare each swimmer to a mental image of distress, but as the beach gets more crowded, you cannot check everyone fast enough. In this paper, the detection pipeline learns what signal patterns look like from examples, so it finds mergers instantly and can estimate how often it might mistake noise for a signal without re-scanning all the data.
As gravitational-wave detectors become more sensitive, they will spot black hole mergers every few minutes. Current search methods are too slow to keep up. This transformer pipeline detects signals quickly and estimates false alarms at a fraction of the cost, making it practical for the high-volume searches expected in the coming decade.
The pipeline is tested only on binary black hole signals within its training range, so performance on other types of gravitational waves or out-of-distribution signals is not established. The false-alarm estimation via time slides may rely on assumptions about noise stationarity. The abstract does not discuss generalization to different detector configurations or noise conditions.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
A Fast and Scalable Transformer Pipeline for Binary Black Hole Detection — Chayan Chatterjee, Abigail Petulante, Haowei Fu et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.