Physics Front Ends with Small Learned Heads for Detecting Dispersed Bursts in CHIME/FRB Baseband Voltages

We evaluate physics-constrained and generic learned detectors of dispersed radio bursts in complex CHIME/FRB baseband voltages, using synthetic bursts injected into real noise and interference from 29 events, with per-event thresholds at a nominal 1% false-alarm rate and seeds chosen on validation events. A hybrid that keeps a bank of 16 coherent dedispersion filters as an analytic complex front end and reads the dedispersed power with a small non-linear head detects 93% of bursts at an energy of 200 noise units with 13k parameters, and 98% with 47k, against 64–73% for generic convolutional networks of 42k–296k parameters, 51% for dedispersion at the true dispersion measure with a boxcar search, and 53% for the same filters with a linear read-out (65 parameters). It reaches the same detection rate at about half the burst energy of the generic networks, holds on the three test events not used during development, and is the only detector that handles the two test events with heavy interference.

On a frozen set of ten real bursts, models with the physics front end localise 5–6, as many as dedispersion at the true DM (6), against 4–5 for generic CNNs trained on the same broad burst family (2–4 on the narrower one) and 1 for an energy detector; the test cannot yet separate the hybrid from the linear filter bank, so its gain on injected bursts is not yet shown on real ones. Matched-size complex and real CNNs perform similarly.

We also find that CHIME baseband channels have the reverse frequency orientation to the textbook dispersion formula, and provide a native implementation of the filter layers in the CNet complex-valued framework with numerical gradient checks and CPU/GPU parity. Separately, an exploratory phase-jump experiment is consistent with a decayed Born-attention streaming model learning the period of the Vela pulsar in real Parkes baseband.

Preprint, October 2026. Data: CHIME/FRB baseband release (doi:10.11570/23.0029) and the Breakthrough Listen open-data archive. Code: the open-source CNet framework.