A Fast and Scalable Transformer Pipeline for Binary Black Hole Detection

Mon, 31 Aug 2026 20:34:23 +0000

Castor, the Coincident Analysis Siamese TransfORmer, is designed for scalable binary black hole searches in Advanced LIGO data. Each detector stream is processed independently using shared transformer weights, and the resulting single-detector outputs are combined into a coincident statistic. This allows time-slide backgrounds to be generated from cached network outputs instead of rerunning the neural network for every slide.


Fig. 1: Castor tokenizes whitened H1 and L1 strain streams, processes them through shared transformer encoders, and combines single-detector log-odds and frame-localization profiles into a post-hoc coincident ranking statistic.

Castor

Mon, 31 Aug 2026 20:34:23 +0000

Castor is a compact transformer-based coincident search pipeline for binary black hole gravitational-wave signals in Advanced LIGO data. It uses a Siamese architecture: the H1 and L1 detector streams are analyzed independently with shared transformer weights, producing per-detector detection statistics and frame-level merger-localization profiles.

The key practical advantage is scalable background estimation. Because Castor caches single-detector outputs, time-slide false-alarm-rate estimates can be generated by recombining those outputs rather than rerunning the neural network for each shift. This makes the pipeline well suited for low-latency and offline searches where robust empirical backgrounds are essential.


Fig. 1: Architecture of Castor. Shared transformer encoders process H1 and L1 strain streams, attention and frame heads produce detector-level summaries, and a post-hoc statistic combines them into a coincident ranking score.

LIGO | Chayan Chatterjee

A Fast and Scalable Transformer Pipeline for Binary Black Hole Detection

Castor