Deep Synoptic Array Science: Searching for Long Duration Radio Transients with the DSA-110

TL;DR

The paper documents the design, commissioning, and early scientific results of the DSA-110 NSFRB project, a GPU-accelerated image-plane single-pulse search pipeline targeting long-duration radio transients in the Galactic Plane. The cerberus pipeline performs real-time de-dispersion and multi-timescale boxcar filtering, with PSF-based candidate ranking and a CNN-based RFI classifier, validated through injection tests and pulsar observations. A pilot Galactic Plane survey yields no detections but establishes a 90% completeness flux limit around 1200 mJy and places Poissonian burst-rate upper limits, informing constraints on White Dwarf-M Dwarf binary and magnetar LPRT models. The work demonstrates feasibility of real-time LPRT searches with a drift-scan radio interferometer and motivates expanded surveys and triggering strategies to improve sensitivity and localization of long-duration radio transients.

Abstract

We describe the design and commissioning tests for the DSA-110 Not-So-Fast Radio Burst (NSFRB) search pipeline, a 1.4GHz image-plane single-pulse search sensitive to 134ms-160.8s radio bursts. Extending the pulse width range of the FRB search by 3 orders of magnitude, the NSFRB search is sensitive to the recently-discovered Galactic Long Period Radio Transients (LPRTs or LPTs). The NSFRB search operates in real-time, utilizing a custom GPU-accelerated search code, \texttt{cerberus}, implemented in Python with JAX. We summarize successful commissioning sensitivity tests with continuum sources and pulsar B0329+54, estimating the 90% completeness $25σ$ flux (fluence) limit to be ~1200mJy (~160Jy ms). Future tests of recovery of longer timescale transients, e.g. CHIME J1634+44, are planned to supplement injection testing and B0329+54 observations. An offline DSA-110 NSFRB Galactic Plane Survey was conducted to search for LPRTs, covering $-3.5^\circ<b<5.7^\circ$ and $141^\circ<l<225^\circ$ (~770 square degrees) in Galactic coordinates. We estimate an upper limit Poissonian burst rate ~2 hr$^{-1}$ per square degree (~17 hr$^{-1}$} per $3^\circ\times3^\circ$ survey grid cell) maximized across the inner $|b|<0.25^\circ$ of the surveyed region. By imposing the ~1200mJy flux limit on two representative models (the magnetar plastic flow model and the White Dwarf-M Dwarf binary model), we reject with 95% confidence the presence of White Dwarf-M Dwarf binary LPRTs (beamed in a detectable direction) with periods between ~10-50s within ~95% of the surveyed region. Combined with the prevalence of LPRTs in the Galactic Plane, our results motivate further consideration of both White Dwarf-M Dwarf binary models and isolated magnetar models. We will continue to explore novel LPRT search strategies during real-time operations, such as triggered periodicity searches and additional targeted surveys.

Figures (11)

• Figure 1: Diagram of the NSFRB Search System. Voltages for $N_{\rm ant}=97$ antennas, $N_t=25$ time samples, $N_p=2$ polarizations, and $N_f=384$ frequency channels are pulled from a rotating PSRDADA buffer, cross-correlated, down-channelized to $n_f=8$, and calibrated. $N_{\rm pix}=301\times301$ images are formed for each of $N_c=16$ sub-bands and sent to the Process Server which assigns search and post-processing tasks. Image data and candidate metadata are saved to disk along with 26.8-second of voltage data triggered from the PSRDADA buffer if candidates above the $6\sigma$ threshold are identified. We outline in green the post-processor which is not required to operate in real-time.
• Figure 2: Pulse width and $6\sigma_{\rm srch}$ Fluence Sensitivity to $\Delta t =134$ ms Bursts of the DSA-110 FRB (light blue; $\sim 840\,$mJy ms) and NSFRB (dark blue hatched; $\sim 40\,$Jy ms) searches. For each search we assume the fluence threshold is fixed; thus the lower border indicates the corresponding flux density threshold $6\sigma_{\rm th}\Delta t/W$ where $W$ is the width measured at 1.4 GHz (note that in some cases li202444wang2025detection, only sub-pulses were detected at 1.4 GHz while a broader profile was detected at lower frequencies. In these cases, we still use the 1.4 GHz sub-pulse width). The dark blue dashed line shows the 90% completeness limit $25\sigma\approx1200$ mJy ($25\sigma\Delta t\approx 160$ Jy ms; see Section \ref{['sec:detrate']}). The purple dashed line shows the theoretical radiometer $6\sigma_{\rm th}$ sensitivity of the NSFRB survey. 1.4 GHz fluences and widths are shown for radio pulsars in the ATNF catalog (diamonds), FRBs from the literature (triangles), and LPRTs (circles). DSA-110 FRBs have dark blue outlines and show their detection pulse widths and S/N rather than refined estimates from baseband analysis. The colorbar indicates distance estimates in kpc (FRB luminosity distances are computed from host Galaxy redshifts; pulsar, magnetar, and LPRT distances are from parallaxes where available, and DM distance limits for the rest of the sample). Note that only pulsars, FRBs, and LPRTs with 1.4 GHz detections are shown.
• Figure 3: Relative S/N Loss (Equation \ref{['eq:snr']}) of NSFRB Search on the Three cerberus Timescales (134 ms, 670 ms, and 10.05 s) as a Function of DM. Each line corresponds to an intrinsic pulse width $W_{\rm int}$, and solid points indicate each trial DM.
• Figure 4: Detection Rate vs. Input S/N for Injection Test. The blue curve includes any candidates detected with S/N$>6\sigma$ while the red curves include only candidates classified as sources by the CNN classifier using probability thresholds $p_{\rm rfi}<50\%$ (dashed) and $p_{\rm rfi}<15\%$ (solid). The black line indicates the $6\sigma$ threshold determined in Section \ref{['sec:initsnr']}.
• Figure 5: Example NSFRB candidate plot from detections of pulsar B0329+54. (top left) Position of the candidate (shaded circle) in reference to the pulsar's catalogued position (open blue square) and other field NVSS continuum sources (open blue circles). The colorbar corresponds to the S/N. (top right) S/N as a function of search DM and pulse width. (middle) DM=0 0.134 ms-sampled time series of the peak candidate in arbitrary units. (bottom) Dynamic spectrum of the peak candidate with coarse 12 MHz channels. The title gives the UTC time of the first time sample, RA and declination after systematic error correction, detected DM, pulse width, and S/N, and the CNN classifier's p-value.