TransientX: Conventional Radio Transient Search

• TransientX is a CPU-based single-pulse and FRB search package that implements a dedispersion workflow across many trial DMs to generate candidate events.
• It serves as a conventional benchmark against which DM-free, multibeam machine-learning pipelines are evaluated in terms of speed and efficiency.
• Integrated in Galactic Centre surveys, TransientX combines traditional RFI mitigation with polarimetric diagnostics to enhance candidate vetting.

Searching arXiv for papers mentioning “TransientX” to ground the article in current literature. TransientX is treated in recent radio-astronomy literature as a conventional transient-search pipeline for radio data and, more specifically, as a CPU-based single-pulse/FRB search package used to find single pulses and fast radio burst candidates by standard dedispersion-based methods. In current arXiv-visible usage, it appears both as an operational search engine in pulsar and transient surveys and as a benchmark representing the traditional DM-grid search paradigm against which a DM-free, multibeam machine-learning pipeline is compared. In that role, TransientX stands for the established workflow in which many trial dispersion measures are searched, data are dedispersed at each trial, signal-to-noise ratio is computed, and candidate events are generated for further inspection (Chen et al., 22 Dec 2025, Desvignes et al., 11 Jul 2025).

1. Definition and place in radio transient search

In the 2025 DM-free FRB study, the software is cited as “\textsc{TransientX}” and the reference is identified as “TransientX: A high-performance single-pulse search package.” In that paper, it is treated as a traditional transient-search pipeline for radio data, designed to find single pulses and FRB candidates by standard radio-astronomy searching methods. The same paper uses it as the main conventional software benchmark, explicitly as a representative CPU-based single-pulse/FRB search package against which the proposed DM-free, multibeam machine-learning pipeline is evaluated (Chen et al., 22 Dec 2025).

This positioning is significant because it places TransientX within the mainstream search stack used for high-rate radio surveys rather than as a specialized post-processing utility. A plausible implication is that, in the literature where it appears, TransientX functions as a reference implementation of the conventional search regime: accurate, operationally relevant, and computationally dominated by dedispersion.

2. Conventional dedispersion workflow

The conventional FRB search workflow associated with TransientX is described in five recurrent stages (Chen et al., 22 Dec 2025):

1. Try many dispersion measures (DMs) over a broad range.
2. Dedisperse the data at each DM trial to align the dispersed burst signal.
3. Compute signal-to-noise ratio (S/N) for each trial.
4. Identify candidate single-pulse events once the optimal DM gives a strong enough detection.
5. Produce a large candidate list for further analysis.

Within this workflow, the most time-consuming step is dedispersion, because the DM is unknown in advance and many trial DMs must be searched. The 2025 FRB paper explicitly groups TransientX with presto and heimdall as standard dedispersion-based tools. Its role is therefore not merely that of one search package among many, but that of a concrete example of the brute-force DM-trial method that newer approaches seek to bypass (Chen et al., 22 Dec 2025).

The computational bottleneck becomes more acute in multibeam observing. The same work argues that, when multibeam receivers are used, conventional dedispersion becomes even more expensive because the data volume scales with the number of beams. This is the immediate context in which TransientX is used as the benchmark for the “old paradigm.”

3. Benchmark role in DM-free multibeam machine learning

A central use of TransientX in recent literature is as the runtime and workflow baseline for a dedispersion-free classifier operating directly on raw multibeam dynamic spectra. The comparison is made against both a multi-beam EfficientNet model and a single-beam ML model, while TransientX is run in single-threaded mode on a benchmark set composed of 7 beam files per case, with 1 to 5 signal sources, for a total of 35 files (Chen et al., 22 Dec 2025).

Benchmark aspect	Reported setting or result
Conventional baseline	single-threaded TransientX
Compared against	multi-beam EfficientNet model; single-beam ML model
Test data	7 beam files; 35 files total
Main speed claim	about 9 times faster
ML throughput	35 data files in 270 seconds; 0.92 TB/hr
Hardware note	Intel Xeon Platinum 8358P (128 cores, 2.6 GHz); NVIDIA GeForce RTX 4090 (24GB)

The paper’s main efficiency claim is that the ML approach is approximately 9 times faster than TransientX, with the stated reason that the model can recognize the multibeam data simultaneously and report all the results together. The optimized EfficientNet-based pipeline is reported to process 35 data files in 270 seconds on a single GPU, corresponding to 0.92 TB/hr throughput. The authors additionally note that the CPU hardware used for TransientX—an Intel Xeon Platinum 8358P processor (128 cores, 2.6 GHz)—costs more than the NVIDIA GeForce RTX 4090 (24GB) used for the machine-learning models, and they use this to argue cost-effectiveness in practical deployment (Chen et al., 22 Dec 2025).

The technical contrast is direct. TransientX embodies a search strategy that requires repeated dedispersion trials and produces candidate lists requiring further filtering, whereas the DM-free model ingests raw multibeam chunks, smooths or resizes them to model input, classifies each chunk directly as FRB or non-FRB, uses multibeam context to suppress RFI, and may estimate DM later with a separate VAE-based method. This suggests that TransientX serves in that paper not only as a speed baseline but as the operational definition of the computation that the new method seeks to eliminate.

4. Use in Galactic Centre transient surveying

TransientX also appears as an operational search tool in a 2025 survey of pulsars and transients in the inner $10\,\mathrm{pc}$ around Sgr A$^*$, based on Effelsberg 100-m telescope observations with a C-X band receiver covering 4 to 9.3 GHz nominally and about $3.4\,\mathrm{GHz}$ effectively after flagging. In that survey, the data were calibrated and cleaned first, and then searched with PulsarX, TransientX and PRESTO. The pipeline had two parallel branches: a periodicity branch based on PulsarX plus PRESTO, and a transient branch in which TransientX was used for all 37 pointings to search for individual pulse-like events (Desvignes et al., 11 Jul 2025).

Several operational details are explicit. The raw data were manually RFI-cleaned using PSRCHIVE; roughly 15% of the bandpass was flagged on average; the data were polarisation-calibrated using noise-diode observations; to avoid requantisation, the calibrated full-Stokes data were stored as 32-bit floats; and channels corrupted in the calibration scan were zero-weighted and discarded later. TransientX was then modified so it could ingest full-Stokes search-mode data and apply rotation-measure synthesis to the detected pulse window, making polarisation part of the candidate inspection products. Output events with $\mathrm{S/N} > 6$ were recorded for visual inspection (Desvignes et al., 11 Jul 2025).

The survey used TransientX in two distinct ways:

1. Standard single-pulse search: this was the normal fast-transient search across all pointings. It produced about 2500 candidates across the 37 pointings. Roughly 10% of them were attributed to the known magnetar SGR J1745–2900, mostly from the inner pointing. No promising single-pulse candidate above $\mathrm{S/N}>8$ was found, except for one tentative narrow pulse at $\mathrm{DM}=4550~\mathrm{pc\,cm^{-3}}$ with $\mathrm{S/N}>7$ (Desvignes et al., 11 Jul 2025).
2. Dedicated search for slow transients in linear polarisation: this was performed only for the inner pointing on Sgr A$^*$. The calibrated PSRFITS data were time-scrunched by a factor of 128, giving a time resolution of about $16.7\,\mathrm{ms}$. The data were dedispersed in Stokes space with DM steps of $200\,\mathrm{pc\,cm^{-3}}$ up to $^*$0. A Faraday correction was applied over

$^*$1

in steps of $^*$2, using

$^*$3

The corrected linear polarisation was frequency-summed and searched with single_pulse_search.py using boxcars up to 300 bins and a minimum detection threshold of $^*$4 (Desvignes et al., 11 Jul 2025).

5. RFI, multibeam context, and polarimetric diagnostics

A major point of comparison between TransientX and newer methods concerns RFI handling. In the multibeam machine-learning work, the impact of radio frequency interference for single-beam and multibeam data is explicitly investigated, and the authors report that ML can naturally mitigate RFI under the multibeam environment. Their stated rationale is structural: RFI usually appears in all beams, whereas an FRB usually appears in only one beam. TransientX is therefore the natural baseline for testing whether multibeam context can replace substantial portions of conventional dedispersion-heavy search (Chen et al., 22 Dec 2025).

In the Galactic Centre survey, TransientX was embedded in a different but related mitigation strategy based on polarisation diagnostics. Polarisation information was used as a scoring input in manual candidate inspection, and for TransientX the candidate plots included the RM synthesis result for the pulse window. The dedicated slow-transient linear-polarisation search was motivated by the claim that single-dish Galactic Centre data are dominated by red noise in total intensity, so standard Stokes $^*$5 searches lose sensitivity to very wide pulses. For sufficiently large $^*$6,

$^*$7

the Faraday correction depolarises baseline fluctuations, making the linear-polarisation search less sensitive to the usual red-noise contamination. At the same time, the survey documents a limitation: after excluding a few problematic two-second windows with quasi-periodic $^*$8 minute artifacts, the search still contained 1033 pulses with $^*$9, attributed to weather radar emitting polarised radiation near $3.4\,\mathrm{GHz}$0 in Germany or to the magnetar. Bright events with $3.4\,\mathrm{GHz}$1 clustered near $3.4\,\mathrm{GHz}$2 and $3.4\,\mathrm{GHz}$3, consistent with SGR J1745–2900 (Desvignes et al., 11 Jul 2025).

A plausible implication is that TransientX, although grouped with conventional dedispersion-based search tools, can be integrated into richer candidate-vetting workflows that include full-Stokes calibration and RM-synthesised diagnostics.

6. Naming ambiguities and unrelated uses

The name “TransientX” should be distinguished from several nearby terms in the arXiv literature.

First, “Transience” in countable Markov decision processes is an objective defined by

$3.4\,\mathrm{GHz}$4

meaning that no state is visited infinitely often. This is a tail objective in a probabilistic verification setting and is unrelated to radio transient searching (Kiefer et al., 2020).

Second, “TRANX” is the TRANSition-based neural abstract syNTaX parser for semantic parsing and code generation. It maps natural-language utterances into formal meaning representations by generating abstract syntax trees from an ASDL grammar, and it is likewise unrelated to radio-search software (Yin et al., 2018).

Third, a separate synthesized usage associated with transaction-oriented networks describes “TransientX-style flattening” as the claim that delayed management costs can be absorbed into immediate operational costs, producing a flattened cost model that preserves resilience behavior. In that account, the core idea is a transformation from a model with operational cost $3.4\,\mathrm{GHz}$5 and long-term impact factor $3.4\,\mathrm{GHz}$6 into a behaviorally equivalent flat model with $3.4\,\mathrm{GHz}$7 and modified upfront cost $3.4\,\mathrm{GHz}$8 (Zinoviev et al., 2014).

In current usage grounded in radio-survey and FRB-search papers, however, TransientX most clearly denotes the conventional single-pulse / transient search package used in dedispersion-based pipelines and as a benchmark against DM-free multibeam classification systems.