Can a GPU-accelerated imaging pipeline enable real-time SETI searches at the Allen Telescope Array?

Imaging replaces beamforming with many lines of sight

Traditional technosignature and transient searches usually work on beamformed time-frequency products. This paper instead builds an interferometric imaging path for Allen Telescope Array (ATA) post-correlation UVH5 visibility data, so that time-domain searches can be run across many image pixels at once rather than extracting a single known position before temporal analysis. The stated motivation is that imaging adds spatial context, which supports localizing candidates and separating celestial emission from terrestrial interference.

The pipeline forms Stokes-I dirty images from correlated visibilities, folds Inter-Quartile Range Mitigation (IQRM) flags into the imaging weights, images 48 frequency subbands spanning 287.5 MHz, subtracts a temporal mean over the full observation from each subband and pixel, applies IQRM-informed subband masking, and performs a blind trial-DM search using a short-history image ring buffer. The dispersion search is performed directly on frequency-resolved images, so each image pixel is searched as an independent line of sight.

The authors frame this as complementary to beamformed backends rather than a replacement: beamformed broadband pulsed searches and real-time deep-learning approaches could provide a first-pass detection stage, with promising candidates then examined using interferometric imaging. The imaging pipeline is intended to support that follow-up by retaining spatial information and searching many pixels simultaneously.

A pulsar observation with a deliberately injected pulse

The ATA was pointed at the bright pulsar PSR J0332+5434 (B0329+54) for a roughly 10-minute observation on 2026 January 21. The observation comprises seven adjacent spectral windows (SPWs) recorded in UVH5 format; each SPW has 192 channels with a 0.5 MHz channel width, giving 96 MHz per SPW and 672 MHz total recorded bandwidth. The dataset contains 6104 integrations at a cadence of about 98.29 ms, and 406 baseline products, of which 378 cross-correlations are used for imaging while the 28 auto-correlations are excluded. Dirty images are formed independently for each integration on a 128x128 pixel grid with 45 arcsec cells.

Early continuum dirty-imaging tests used a single SPW (SPW #3), which had relatively little RFI, and the pulsar appears periodically in that image time series, consistent with its roughly 0.7145 s pulse period. Because the pulsar's dispersion measure is low, the in-band dispersive sweep across a single 96 MHz SPW is only a small fraction of an integration and does not strongly test image-domain DM search. The multi-band results therefore use the first three SPWs and inject a bright, strongly dispersed pulsed signal.

Those three SPWs provide 576 contiguous frequency channels spanning 287.5 MHz, from 1.164 to 1.452 GHz, on the same time and baseline grid; the remaining four SPWs are reserved for future wider-bandwidth tests. The injected source is a periodic dispersed pulsed signal with DM = 150 pc cm-3, added directly in the visibility domain as a real-valued complex constant of amplitude 10^6 in native, uncalibrated visibility units on every baseline product, with zero imaginary part. No target input S/N is specified. Because the added phasor is identical across baselines, the dirty-image response peaks at the phase centre, pixel (64,64) on the 128x128 grid. Each pulse has an intrinsic width of one integration, about 98.29 ms, and pulses are injected every 100 integrations, a period of about 9.829 s.

Image-plane S/N and temporal S/N answer different questions

The paper reports two S/N values that readers should not treat as the same measurement. After image-domain dedispersion, the injected pulses have a median spatial image-plane S/N of about 52, measured relative to the spatial RMS of source-free pixels in the corresponding DM = 150 pc cm-3 dedispersed images. This compares the source with the spatial noise in an individual dedispersed image.

The blind search, by contrast, ranks candidates using a temporal statistic computed from the full dedispersed time series at each pixel and DM, with S/N defined as the maximum minus the median divided by 1.4826 times the median absolute deviation. On the three-SPW test set, the blind search recovers the injection at the phase centre with best-fit DM = 150 pc cm-3 and a temporal MAD S/N of about 286; neighbouring trials at 140 and 160 pc cm-3 give about 272. This temporal value compares the strongest pulse with the variation over time at the same pixel, and the ranking selects the strongest individual pulse in the injected periodic train.

The distinction matters because the two numbers describe different denominators and different questions: one is a per-image spatial contrast, the other is a per-pixel temporal contrast across the observation. A reader comparing this pipeline with a beamformed search should check which statistic is being quoted before drawing conclusions about relative sensitivity.

Delays are computed per channel, not per subband

Frequency-dependent arrival times follow the cold-plasma dispersion law. A delay is computed independently for each of the 576 native frequency channels, not per subband, and every channel frequency is compared to the same global reference: the highest channel centre across all three SPWs, 1.4515 GHz. Lower-frequency channels are therefore delayed relative to that reference. Delays are evaluated once on the concatenated 576-channel frequency axis and then applied within each SPW file, so the three bands share a common dispersion reference rather than independent per-file references.

Continuous delays are rounded to the nearest integer number of integrations before injection. For DM = 150 pc cm-3, the largest delay is that of the lowest channel centre, 1.164 GHz, at about 0.164 s, or roughly 2 integrations.

For the blind search, trial dispersion measures are DM = 10, 20, ..., 200 pc cm-3, giving 20 trials. Subband delays use the same cold-plasma relation but are evaluated at each subband centre frequency rather than at the native channel centres used for injection, and continuous delays are again rounded to the nearest integer number of integrations. After applying these integer time shifts, the subband images are averaged over the usable subbands at each pixel, ignoring NaN-valued planes, to form the dedispersion cube.

The same RFI flags are used two different ways

RFI flagging follows IQRM, which identifies frequency channels whose visibility statistics are anomalous relative to neighbouring channels. Mask generation is performed once, offline, on the concatenated three-SPW visibility set: form Stokes I; build a time-frequency dynamic spectrum by incoherently summing absolute visibility amplitudes across cross-baselines; concatenate the three SPWs in frequency; within blocks of 1000 integrations reduce each block to a per-channel spectral statistic, here the temporal standard deviation, and apply IQRM neighbour comparisons with threshold 3.0 and a radius of order 10% of the channel count (radius 58 for 576 channels); then produce a boolean mask. On the three-SPW test set, the joint mask occupancy is approximately 29.1% of time-frequency samples, and 246 of 576 channels are flagged at least once.

During imaging, individual flagged channels are assigned zero weight across all baselines for that integration, while unflagged channels in the same subband still contribute to the dirty image. The visibility amplitudes are not rewritten and the UVH5 files are not modified. For the DM search, the authors adopt a more conservative criterion: if any of the 12 channels in a subband is flagged at a given integration, the entire 128x128 subband plane is excluded from the trial-DM combination at that integration. In this dataset, 20 of the 48 subbands are excluded at every integration.

The authors compared this conservative rule with two less aggressive alternatives: keeping a subband whenever at least one of its 12 channels was unflagged, and weighting each subband by the fraction of unflagged channels. Both alternatives retain much more data, about 89.7% of subband-time samples versus about 50.3% for the conservative rule, but on this dataset they let RFI dominate the blind search and obscure the injected pulse. The conservative any-channel rejection is therefore used for all reported results, and the authors note that milder or spatially dependent RFI treatments remain to be explored.

A short image history avoids storing the full cube

Storing the full subband cube of 48 x 6104 x 128 x 128 float32 values would require approximately 17.4 GB. For the present DM grid, however, the maximum integer delay is only 2 integrations, so only a short history of recent subband images is needed. The pipeline uses a ring buffer of depth R = 8, which satisfies R greater than the maximum delay plus one and is sufficient to form all trial-DM alignments without retaining the full frequency-by-time-by-image cube in memory. This reduces the resident image history to about 30 MB.

The trial-DM image combination was about 5.2 times faster when operating on the ring buffer than when scanning the full stored cube. The authors state that the buffering strategy is inspired by the streaming image-set and ring-buffer concepts in STRIDE, but that here the ring serves as a short-history buffer for forming dispersion-aligned image combinations and is not a full implementation of STRIDE.

Core science fits the cadence, output writing does not

When integrated into the CyberEther streaming framework, the core science stages require approximately 63.3 ms per integration, below the native cadence of about 98.29 ms. The per-stage breakdown in the paper is roughly 9.66 ms for replay and I/O, 0.09 ms for IQRM weight application, 33.05 ms for GPU subband imaging, 0.18 ms for mean subtraction, 0.18 ms for subband masking, 0.10 ms for ring buffer management, and 20.08 ms for CPU trial-DM combination. Only dirty imaging runs on the GPU; replay, IQRM weight application, mean subtraction, IQRM-informed subband masking, ring management, and trial-DM combination remain CPU stages in the present pipeline.

Writing the full dedispersed product adds approximately 50 to 55 ms per integration, increasing the total processing time to about 124 ms per integration. The authors therefore distinguish the science processing time from output I/O and explicitly do not treat the I/O-inclusive run as real-time. This is the key threshold in the paper: the science path fits inside one integration, but the complete pipeline as currently configured does not.

The paper also reports that on a single 96 MHz SPW, end-to-end continuum imaging, one dirty image per integration including display mapping, takes about 2.9 ms per integration versus the native 98 ms cadence. That figure applies to the single-SPW continuum imaging path, not to the full three-SPW image-domain DM search, and should not be substituted for the 63.3 ms science total.

Recorded replay is not yet a live correlator backend

The present implementation streams recorded post-correlation visibility data through the pipeline rather than operating as a live ATA correlator backend. It relies on a temporal mean and a MAD noise estimate precomputed over the full observation. For live operation, these quantities will need to be estimated from data available up to the current integration. Candidate ranking is currently performed offline after replay rather than during streaming, and periodicity is not included in the present candidate ranking stage, although the authors have developed tools that can browse dedispersed time series and run FFT and fold follow-ups.

The authors also note that GPU subband imaging is divided into 48 relatively small operations with repeated data transfers, so GPU utilization is not yet optimal. Evaluation so far has emphasized a controlled injected signal rather than blind searches on real data. The conservative any-channel subband mask worked well on this dataset, but milder or spatially dependent RFI treatments remain to be explored.

Planned work includes batching GPU subband imaging to improve occupancy beyond the present 48 serial subband operations, running background and noise estimators for live processing, performing candidate detection during streaming without large dedispersed data products, and periodicity scoring on dedispersed per-pixel time series. The authors also plan to test the full image-domain DM search on individual pulses from known, more highly dispersed pulsars, including B1933+16 at DM around 158 pc cm-3, and to run injection tests spanning a range of DMs and S/N ratios to characterize recovery performance.

A validated building block, not a finished survey system

The paper's specific claim is narrow and testable: on a controlled three-SPW injected dataset, the pipeline recovers the dispersed pulsed source at the phase centre with DM = 150 pc cm-3 and temporal S/N of about 286, and the core processing finishes within one integration at about 63 ms versus the roughly 98 ms cadence. The authors are explicit that writing the large dedispersed product does not fit that budget.

For readers weighing this against beamformed search backends, the decision-relevant comparison is not raw sensitivity but where the pipeline sits in a survey chain. The paper envisions pairing beamformed first-pass searches with this imaging pipeline, so that candidate events identified in beamformed products could trigger extraction of the corresponding visibility or image data segment for image-domain localization and confirmation. The value demonstrated here is that the imaging and image-domain DM search stages can run within the ATA integration time on recorded data, with the remaining gaps being live background and noise estimation, streaming candidate ranking, and output I/O.

The most useful way to read the result is therefore as a validated building block with a measured timing budget, not as a completed real-time SETI system. The 63.3 ms science total and the roughly 124 ms I/O-inclusive total are the two numbers that define how far the current implementation has moved toward live operation, and the conservative subband masking result, which retained only about 50.3% of subband-time samples but was necessary to keep RFI from obscuring the injected pulse, is the paper-specific trade-off that any follow-up implementation will have to revisit on real data.

Sources used in this answer