The Northern High Time Resolution Universe pulsar survey: II. Single-pulse search set-up and simulations

The High Time Resolution Universe (HTRU) survey is an all-sky survey looking for pulsars and other radio transients. A new single-pulse (SP) search pipeline is presented, tailored to the northern part of the HTRU survey collected with the 100m Effelsberg Radio Telescope. In a selection of this data, synthetic SPs are injected with frequency-time structures resembling those of the detected Fast Radio Burst (FRB) population and processed by the pipeline to characterize its performance. Therefore, several new software toolkits have been developed (FRBfaker and RFIbye) to enable the injection of SPs with complex frequency-time structures and cope with the Radio Frequency Interference (RFI) in the survey’s data. The operation of these toolkits is described alongside the overall functionality of the SP pipeline. Qualification of the pipeline confirmed that it is ready to process all the HTRU-North data. Additionally, the survey’s sensitivity to SPs, the impact of RFI thereon, the performance of the deep-learning classifier FETCH, and some insights that may be used to improve the pipeline’s performance in the future are determined. Within the small data sample analysed, 21 known pulsars and a RRAT are detected. In addition, eight faint SP trains that might originate from yet undiscovered neutron stars and 141 isolated SP candidates were discovered.

💡 Research Summary

**
The paper presents a dedicated single‑pulse (SP) search pipeline designed for the northern component of the High Time Resolution Universe (HTRU‑North) survey, which is conducted with the 100 m Effelsberg radio telescope. The authors describe the observational setup, data characteristics, and the selection of a representative data subset (1 500 mid‑latitude pointings, of which 1 000 are used for performance testing). Each pointing provides filterbank files with 512 frequency channels, a 300 MHz bandwidth centred at 1360 MHz, and a time resolution of 54.61 µs.

The pipeline is built to run on a GPU‑enabled super‑computing cluster and consists of five main stages: (1) data retrieval and integrity checking; (2) an initial search for SP candidates using the GPU‑based software heimdall, with a lowered DM tolerance (1.05) and a signal‑to‑noise (S/N) threshold of 6.5, covering dispersion measures from 10 to 5 000 pc cm⁻³ and pulse widths from 0.109 to 13.98 ms; (3) radio‑frequency‑interference (RFI) mitigation via a custom Python script called RFIbye, which replaces RFI‑contaminated samples with statistically similar noise on a channel‑by‑channel basis; (4) a second heimdall run on the cleaned data followed by a set of heuristic filters (minimum of three boxcar/DM trials, width larger than the expected DM smearing) to produce a list of “valid candidates”; and (5) candidate classification using the deep‑learning framework FETCH, which evaluates each candidate with eleven pre‑trained neural‑network models and returns an averaged probability of being a genuine astrophysical transient. FETCH is used only as a decision‑support tool rather than a hard pre‑selection filter, preserving the possibility of discovering non‑FRB‑like phenomena.

To quantify pipeline performance, the authors developed two auxiliary toolkits: FRBfaker, which injects synthetic single pulses with realistic frequency‑time structures (including spectral indices, scattering tails, and fluence distributions) mimicking the observed FRB population; and RFIbye, which provides a flexible threshold‑based RFI excision scheme. By injecting thousands of synthetic pulses across a range of fluences, DMs, and spectral shapes, they measured a detection efficiency exceeding 90 % for pulses with fluences greater than ~0.2 Jy ms and DMs above ~200 pc cm⁻³. The RFI mitigation step reduced the number of false positives by roughly 70 % without significantly degrading sensitivity.

Applying the fully qualified pipeline to the 1 000 analysis pointings yielded the re‑detection of 21 known pulsars and one rotating radio transient (RRAT), confirming the pipeline’s correctness. In addition, eight faint SP “trains” (multiple pulses from the same sky location) and 141 isolated SP candidates were identified. Most of the new detections lie near the S/N threshold (6.5–8), a regime where traditional searches often miss signals. Candidates appearing in multiple non‑adjacent beams within a short time window were flagged as RFI, while those surviving all automated checks were inspected manually; those with FETCH‑averaged probabilities above 0.5 were highlighted for further follow‑up.

The authors conclude that the pipeline is ready for full‑scale processing of the entire HTRU‑North survey, which will eventually cover ~218 000 pointings. They outline future improvements, including real‑time RFI monitoring, adaptive DM step optimization, incorporation of additional machine‑learning classifiers tailored to non‑FRB transients, and further GPU‑based parallelization to accelerate processing. The work demonstrates that a carefully tuned combination of high‑resolution data handling, sophisticated RFI excision, and deep‑learning classification can significantly enhance the discovery potential for both known and novel radio transients in large‑scale pulsar surveys.