LeoNet: Neural network detects binary pulsars about 500 times faster than conventional search

Ten learned filters replace the tested acceleration grid one by one. On simulated observations, LeoNet misses an average of 55.5% fewer signals than the evaluated PRESTO configuration.

Binary pulsars are valuable laboratories for testing theories of gravitation, but their orbital motion shifts the frequency of their pulses by Doppler effect, which complicates their detection. The reference method consists of testing one by one numerous acceleration and "jerk" hypotheses in the Fourier domain, at the cost of a very cumbersome calculation.

The authors of LeoNet, presented on arXiv, replace this grid with a convolutional neural network with ten learned filters, which directly extract signatures from Doppler-distorted signals. The resulting compact feature map is then examined by a classifier. On simulated 500-second observations, the false negative rate drops by an average of 55.5% compared to the PRESTO configuration evaluated, over five sampling intervals.

A few milliseconds per observation

Optimized with TensorRT, LeoNet processes a complete observation in 3.4 to 4.4 milliseconds on an NVIDIA H100 GPU, including preprocessing. At 128 microseconds, it needs 3.54 ms compared to 1.767 s for the PRESTO search on a processor, a gain of about 499 times. This is enough to consider sorting candidates in real time in pulsar search chains, subject to confirmation on real data.