2.1. Science goals and motivation

The broader goals of the SMART survey are similar to most other large sky surveys, that is, exploring the new parameter space that is opened by a leap in instrumentation, technology, or sensitivity and to uncover a large population of previously undetected pulsars. The fact that the currently known pulsar population ( $\sim$ 3300, cf. the Australia Telescope National Facility (ATNF) pulsar catalogueFootnote a v1.67; Manchester et al. Reference Manchester, Hobbs, Teoh and Hobbs2005) represents only a small fraction ( $\lesssim$ 10%) of the total expected (i.e. beamed in our direction) Galactic population (e.g. Keane et al. Reference Keane2015, and references therein) strongly motivates such large sky surveys. Indeed, conducting a full Galactic census of pulsars is a high-priority science objective for the SKA. Further, given the number of broader questions surrounding the neutron-star population (e.g. birth rates, and comparison with rates of supernovae), the detectable pulsar population is largely guided by the known population of pulsars at any given time. It is therefore imperative to explore every possible avenue and steadily refine our knowledge of pulsar population. Furthermore, the detection prospects of pulsars in a given frequency band strongly depends on the emission and propagation properties at those frequencies; however, the current forecast of a detectable population in the SKA-Low band is largely guided by the pulsar population uncovered by high-frequency surveys.

Table 1.

Parameters of large pulsar surveys over the past decade.

Notes: Survey description reference – SCB+19: Sanidas et al. (Reference Sanidas2019) for LOTAAS; SLR+14: Stovall et al. (Reference Stovall2014) for GBNCC; BCM+16: Bhattacharyya et al. (Reference Bhattacharyya2016) for GHRSS; KJvS+10: Keith et al. (Reference Keith2010) for HTRU; HWW+21: Han et al. (Reference Han2021) for GPPS.

Obtaining a large body of measurements such as DM, scattering and Faraday rotation, by using pulsars as probes of the ISM, will also enable mapping out the distribution of magneto-ionic (and turbulent) plasma in the Galaxy, which is steadily refined with a larger sample of measurements (e.g. Cordes & Lazio Reference Cordes and Lazio2002; Bhat et al. Reference Bhat, Cordes, Camilo, Nice and Lorimer2004; Deller et al. Reference Deller2016; Yao, Manchester, & Wang Reference Yao, Manchester and Wang2017).

Finally, an underlying goal of any large sky pulsar survey is to discover exotic objects; while it is hard to design any particular survey specifically for this, historical examples are abundant, for example, the discovery of the double pulsar in the Parkes multibeam (PMB) survey (Lyne et al. Reference Lyne2004), the 23.5-s period pulsar in LOTAAS (Tan et al. Reference Tan2018b), and the transitional millisecond pulsar (MSP) in the Arecibo drift-scan survey (Archibald et al. Reference Archibald2009). All such broader and high-profile objectives are certainly applicable for the SMART survey.

The SMART pulsar survey also perfectly complements ongoing northern-sky surveys in sky and frequency coverage (Table 1). Surveys at low frequencies will likely be sensitive to a different pulsar population, and therefore an all-sky survey at low frequencies is also essential to develop a comprehensive picture of neutron-star/pulsar populations in the Galaxy. Bearing this in mind (and as we detail in Section 2.4), the survey is designed to reach a final sensitivity comparable to that of LOTAAS, that is, the use of long dwell times (4800 s) to attain a limiting sensitivity (10 $\sigma$ ) of $\sim$ 2–3 mJy for long-period pulsars with small duty cycles, and assuming a spectral index $\unicode{x03B1} = -1.5$ and no turnover down to $\sim$ 150 MHz. This is $\sim$ 3–5 times deeper than the previous-generation low-frequency (70 cm) survey (Manchester et al. Reference Manchester1996) in the south (and thence an accessible search volume $\sim$ 5–10 times larger), and $\sim$ 2–3 times deeper than the high-latitude segment of the Parkes HTRU survey (Keith et al. Reference Keith2010).

The SMART survey will also serve as a reference survey for future deeper surveys at low frequencies, such as those planned with SKA-Low (Keane et al. Reference Keane2015). While the success of (and the lessons learned from) all ongoing low-frequency surveys will indeed inform SKA-Low pulsar surveys, the SMART survey will potentially play an additional important role, since the MWA is also the official low-frequency precursor for SKA-Low, and is located at the same site where SKA-Low will be built. Specifically, the sky coverage of the SMART survey is identical to that of SKA-Low, which means a higher degree of synergistic overlap in calibration and beamforming methodologies, than most northern facilities. The role of reference surveys is vividly demonstrated by the later generation multibeam surveys in the south; for example, the PMB survey for its successors, the HTRU pulsar survey (Keith et al. Reference Keith2010) and the SUrvey for Pulsars and Extragalactic Radio Bursts (SUPERB; Keane et al. Reference Keane2018), which can now play a similar role for the planned surveys with MeerKAT. However, aside from the Parkes 70 cm survey of the 1990s, the low-frequency southern-sky remains essentially unexplored for pulsar searches, especially at $\lesssim$ 300 MHz.

Aside from the aforementioned primary science goals, there are also some auxiliary goals for the SMART survey, largely enabled by the novelty of the data recording strategy, that is, the use of voltage capture system and post-processing, as opposed to the beamformed data in the filterbank format. These not only facilitate a number of additional strategies for confirmation and follow-up, but they can also be potentially exploited for developing and trialling alternate strategies for pulsar searches; for example, image-based techniques for the identification of promising candidates that take advantage of pulsar properties such as steep spectrum, variability or circular polarisation (e.g. Sett et al. Reference Sett, Bhat, Sokolowski and Lenc2022). These, in principle, also offer some advantages over traditional search methods, especially for extreme pulsars like those with sub-millisecond periods, or distant pulsars whose pulse shapes will be significantly broadened due to multi-path scattering, but will be sensitive primarily to very bright sources.

Notwithstanding the anticipated scientific merits of the SMART survey, computational requirements are substantial, especially given the large data rate of the VCS and searching at low frequencies, thereby necessitating a multi-pass processing strategy. In the first-pass processing, we perform a shallow survey, where 10 min of data from each observation are processed, and the search is limited to basic periodicity, and DMs up to 250 ${\textrm{pc cm}^{-3}}$ . In this paper, we outline the observing strategies employed for the survey, and processing strategies adopted for the initial phase, and present analysis and results to date, as well as plans and strategies for future processing. A companion paper (hereafter Paper II) will describe the survey status, pulsar census to date and more details on follow-up strategies including timing and imaging follow-ups.

2.2. Survey strategy

The novel strategy employed for the SMART survey, that is, the use of VCS recording from 128 tiles, which allows high-time resolution (and instantaneous) sampling of a very large patch of the sky (but at the expense of a large data rate of 28 ${\textrm{TBh}^{-1}}$ ), necessitates substantial processing to enable large-scale pulsar searching applications. Most importantly, the voltage data from the tiles need to be coherently combined to generate thousands of tied-array beams prior to any search processing. The undertaking of the SMART survey is particularly enabled by the Phase II upgrade, whereby a compact configuration of 128 tiles within $\sim$ 300 m became available on a semi-regular basis. The compact configuration of Phase II brings an enormous efficiency in terms of beamforming cost; specifically, the number of tied-array (i.e. phased array) beams required to fill the full FoV (at a gain level down to half power point) is reduced from $2.7 \times 10^5$ for the Phase I array to $3.9 \times 10^3$ for the Phase II compact array. This reduction of more than two orders of magnitude in the computational cost makes an all-sky high-sensitivity pulsar search tractable (and affordable) with an interferometric array like the MWA. Thus, with the beamforming step integrated into software-defined instrumentation, this effectively translates into an impressively large survey speed of $\sim 450\, {\textrm{deg}^2}{\textrm{h}^{-1}}$ , that is, the full visible sky of the MWA ( $\unicode{x03B4} < +30^{\circ}$ ) can be surveyed in a modest number of VCS pointings.

The first-pass survey strategy of processing only 10 min of data from each observation (hence reaching about one-third of the full-search sensitivity) was adopted also to boost the prospects of early pulsar discoveries. Even though the combination of the VCS mode and the FoV provides a large survey speed, practical considerations such as the availability of the compact configuration necessitated multiple observing campaigns to advance the survey. Further details including the survey status and completion plans are described in Paper II.

2.3. Beamforming and Sky tessellation

The signal processing chain of the MWA including the high time resolution system is described in a number of earlier papers (e.g. Tingay et al. Reference Tingay2013; Prabu et al. Reference Prabu2015; Tremblay et al. Reference Tremblay2015), and is briefly reiterated here. In the legacy system that was employed for survey campaigns to date, the VCS sub-system follows the second stage of channelisation in the signal path. Each element of the array is a $4 \times 4$ dipole array, called a ‘tile’, the signals from which are fed to an analogue beamformer that defines the FoV. The beamformed signals are Nyquist-sampled at 655.56 Msps and channelised (after signal conditioning) using a polyphase filterbank (PFB) to generate 256 $\times$ 1.28-MHz signal outputs (i.e. coarse channelisation), 24 of which are transported to the central processing facility, where a second-stage PFB operation is performed, resulting in 128 $\times$ 10-kHz time series for each coarse channel, that is, 3072 channels across the recording 30.72 MHz bandwidth. These voltage time series are written to an array of RAID disks by the VCS as 4+4-bit complex voltage samples. These data are recorded (up to a maximum duration of 100 min) and transported to the Pawsey Supercomputing Centre where further processing (including calibration and beamforming) is carried out.

VCS-recorded data can be processed offline for calibration and tied-array beamforming (Ord et al. Reference Ord2019) and, optionally, can also be reprocessed to reconstruct a higher time resolution voltage data at the native coarse-channel resolution of 0.78 $\unicode{x03BC} {\rm{s}}$ (McSweeney et al. Reference McSweeney2020). To realise the SMART pulsar survey, this beamformer functionality was further enhanced to optionally generate several dozens of tied-array beamformed outputs simultaneously—i.e. the so-called multi-pixel beamformer, which is essentially the front end of the pulsar search processing chain. The implementation details and benchmarks are described in Swainston et al. (Reference Swainston2022). This software tied-array beamformer has been benchmarked on Pawsey’s Garrawarla and Swinburne’s OzSTAR supercomputers. It performs $3\times$ faster on the latter, which has been the primary high-performance computing (HPC) platform for much of our SMART data processing.

Thanks to the large FoV of the MWA ( $\sim 610\, {\textrm{deg}^{2}}$ at 155 MHz, near zenith), the entire sky south of declination $\unicode{x03B4} < +30^{\circ}$ can be covered in a modest number of telescope pointings. The sky tiling strategy is shown in Figure 1. In short, we adopted pointings similar to that of the GaLactic and Extragalactic All-sky MWA survey (Wayth et al. Reference Wayth2015), that is, meridian drift scans optimised for maximum sensitivity at each declination as well as for more reliable calibration (referred to as ‘sweet spots’). In this case, the number of pointings depends on the degree of overlap in right ascension (RA), with a minimum of 58 pointings for minimal ( $1^{\circ}$ ) overlap and 78 for a $15^{\circ}$ overlap. A large overlap is more optimal as it effectively serves as a two-pass strategy, which is desirable at low frequencies where intermittency (from effects such as scintillation) tends to be more pronounced. After exploring the full range of options, and also factoring in the available resource constraints, we converged on a $10^{\circ}$ overlap as an acceptable choice.Footnote b As shown in Figure 1, this amounts to a total 70 pointings, that is, 93 h of telescope time for the full SMART survey.

Figure 1.

Sky tessellation of the SMART survey. The left panels show beam tiling patterns for two select pointings: top one a near-zenith pointing ( $\unicode{x03B4}=-28^{\circ}$ ), the bottom one a far southern pointing ( $\unicode{x03B4} = -70^{\circ}$ ). The number of tied-array beams vary from $\sim$ 6000 to $\sim$ 8000 from near-zenith to far-zenith pointings, and the beam shape becomes elliptical at large offsets from the zenith. The size of the circle/ellipse indicates half power tied-array beam size; the red and blue circles correspond to the low and high ends of the SMART band (140–170 MHz). The right panels show the primary beam response for the same declination pointings, at the central frequency of 155 MHz.

Figure 2.

Left: Minimum detectable flux density, ${S_{\textrm{min}}}$ , for the first-pass processing of the SMART survey as a function of DM. Sensitivity limits, assuming a 10-min integration time, are plotted for different pulse periods, $P=$ 1.0, 0.1, 0.01, 0.001 s, and for two different system temperature values ${T_{\rm{sys}}}$ ; one corresponding to mean ${T_{\textrm{sky}}}$ for regions away from the Galactic plane, and the other for a mean ${T_{\textrm{sky}}}$ in the plane, but excluding the region toward the Galactic Centre. The effect of pulse broadening due to interstellar scattering (Bhat et al. Reference Bhat, Cordes, Camilo, Nice and Lorimer2004) is shown by the dotted lines. Right: Pulse broadening (smearing) incurred by using the first-pass processing dedispersion plan (Table 2) due to various factors such as the finite sampling time, dispersive smearing due to the incoherent de-dispersion algorithm used, and the effects of multi-path scattering based on the $\tau_d$ -DM relation from Bhat et al. (Reference Bhat, Cordes, Camilo, Nice and Lorimer2004). The grey shaded region denotes one order of magnitude larger or smaller range in the predicted scattering.

For each pointing, many tied-array beams (TABs) are formed to maximise the sensitivity across the FoV. The tied-array beams are pointed towards fixed right ascension and declination, with the necessary adjustments to the tile phases made for every second of data (Ord et al. Reference Ord2019). Thus, although the observations themselves are drift scans, sources can be tracked by the same TAB for up to the full duration of the observation.

The precise size and shape of the TABs is a non-trivial function of the tile layout of the compact configuration, equivalent to the ‘Compact robust 1’ synthesised beam whose cross section is presented in Figure 7 of Wayth et al. (Reference Wayth2018) and discussed in Swainston et al. (Reference Swainston2022) and in Section 3. Due to the compact configuration’s redundant baselines (in the two ‘hexes’), the most sensitive parts of the TAB consist of a main lobe whose full width half maximum (FWHM) at 155 MHz is 23 $^{\prime}$ , surrounded by a pattern of discrete grating lobes of similar width. Although these grating lobes can be exploited for candidate confirmation (further discussed below), we choose the TAB pointings to form a dense (hexagonal) grid such that the main lobes overlap by $\sim$ 20%, as shown in Figure 1. This effectively Nyquist-samples the sky at a gain of the half-power level or more. The beam shape used for this calculation assumes that all 128 tiles are functioning, whereas, in reality, up to $\sim$ 10% of tiles may be flagged in any given observation. Unless the flagged tiles preferentially result in a reduction of the longest baselines, the effect on the beam shape is negligible.

Tiling the FoV in this way translates to $\sim$ 6300 TABs for an observation pointed toward the zenith. For pointings away from the zenith, where the beam shape develops a significant ellipticity (e.g. at zenith angle 15 $^{\circ}$ , ellipticity $\unicode{x03B5} = \unicode{x03B8} _{\rm maj} / \unicode{x03B8} _{\min} =$ 1.36 where $\unicode{x03B8}_{\rm maj}$ and $\unicode{x03B8} _{\min}$ are the major and minor axes of the TAB), the number of TAB pointings are in the range $\sim$ 4200–4500. Further, the beam size also varies across the 20% fractional bandwidth of our survey observations; for example, for a pointing toward the zenith (where the TAB is nearly symmetrical), the FWHM is 25.3 $^{\prime}$ at 140 MHz but reduces to 20.7 $^{\prime}$ at 171 MHz. This further justifies our rationale for a 20% overlap, as it ensures every single spot in the sky is covered at a gain near or above the half power level even at the high end of the observing band.

Finally, as with any other aperture array, the sensitivity is not uniform across the sky and is strongly declination-dependent; to first order, the loss in sensitivity is by a factor $\cos\!(\unicode{x03B8} _z)$ where $\unicode{x03B8} _z$ is the zenith angle. In principle, this can be compensated to a certain extent by longer integrations, though in practice, the inherent limitations of our data recording system (VCS) limits this to no more than 90 min, and we therefore use 80 min recordings for all pointings. As such, the sensitivity will not be uniform across the sky due to other factors; for example, the sky background temperature ${T_{\textrm{sky}}}$ is direction dependent, and the loss in sensitivity from severe pulse broadening for distant pulsars, which applies to the sight lines within the Galactic plane or toward the Galactic centre. Some of these are considered in detail in Section 2.4.

2.4. Survey sensitivity

The sensitivity of a pulsar survey is determined by the combination of some instrumental and processing parameters and a variety of broadening effects to pulsar signals. Following Dewey et al. (Reference Dewey, Taylor, Weisberg and Stokes1985), the minimum detectable flux density for a pulsar with period P and effective pulse width ${W_{\rm{eff}}}$ , down to detection significance $({\rm S/N}) _{\rm min} $ , that is, minimum detectable signal-to-noise ratio, is related to the telescope gain G and system temperature ${T_{\rm{sys}}}$ , which is the sum of the receiver and sky background temperatures, that is,

(1) \begin{equation} S_{\textrm{min}} = \frac{ ({\rm S/N})_{\rm min} ( T_{\textrm{recv}} + T_{\textrm{sky}} )}{ G \sqrt{ n_{\textrm{pol}} t_{\textrm{obs}} B_{\textrm{obs}} } } \sqrt { \frac{ W_{\textrm{eff}}}{ P - W_{\textrm{eff}} } }\end{equation}

where ${n_{\rm{pol}}}$ is the number of polarisations summed, ${t_{\rm{obs}}}$ is the integration time and ${B_{\rm{obs}}}$ is the recording bandwidth. As evident from this equation, the sensitivity is maximum for long-period pulsars with a small duty cycle, i.e. when $ W_{\textrm{eff}} \ll P $ . The gain $ G = A_{\textrm{eff}} / 2 k_B $ , where ${A_{\rm{eff}}}$ is the effective collecting area and $k_B$ is the Boltzmann constant. At 150 MHz, $ A_{\textrm{eff}} \approx 2750 $ ${{\rm{m}}^2}$ for a 128-tile MWA (Tingay et al. Reference Tingay2013), which may imply $ G \sim 1 $ ${\textrm{K Jy}^{-1}}$ , however, for an aperture array such as the MWA, it is a strong function of the zenith angle, i.e. $G(\unicode{x03B8} _z) = G_{\rm max} {\rm cos}(\unicode{x03B8} _z)$ , where $\unicode{x03B8} _z$ is the zenith angle and $G_{\rm max}$ is the gain at $\unicode{x03B8}_z=0$ . Moreover, for drift-scan type observations that we employ for the SMART, G depends on the offset from the phase centre, and can be $\sim$ 0.5 $G_{\rm max}$ at the half power point. We therefore assume a conservative $G \sim 0.5$ ${\textrm{K Jy}^{-1}}$ for all our sensitivity calculations. This is assuming a full coherent beam sensitivity, that is, perfect calibration for TAB formation and no loss of sensitivity due to flagged tiles. In practice, a small number of tiles ( $\lesssim$ 10) are typically flagged due to malfunctioning, sub-optimal performance or poor calibration solution. As we detail in Section 3, the strategy of observing multiple calibrators for SMART observation allows us to perform useful cross-checks and maximise the achievable sensitivity using the best available calibration solutions.

At the low frequencies of the MWA, the system temperature ${T_{\rm{sys}}}$ is dominated by the sky background ${T_{\textrm{sky}}}$ . Both ${T_{\rm{recv}}}$ and ${T_{\textrm{sky}}}$ are frequency-dependent, and ${T_{\textrm{sky}}}$ is also a strong function of the direction (l, b), where l and b are the Galactic longitude and latitude, respectively. We assume a mean ${T_{\rm{recv}}}$ = 50 K for the 140–170 MHz band. Excluding a $\sim$ 10 $^{\circ}$ cone around the Galactic centre, ${T_{\textrm{sky}}}$ can vary from $\sim$ 200 K toward $|b| \gtrsim 60^{\circ}$ to as much as $\sim$ 1200 K in the plane, toward $\gtrsim$ 10 $^{\circ}$ from the Galactic centre, where ${T_{\textrm{sky}}}$ can be as large as $\sim$ $10^4 $ K at 155 MHz. We use the Haslam et al. (Reference Haslam and Salter1982) map as the reference and assume $ T_{\textrm{sky}} \propto \nu^{-2.55} $ scaling from Lawson et al. (Reference Lawson, Mayer, Osborne and Parkinson1987). Given this strong dependence of ${T_{\textrm{sky}}}$ with (l, b), we consider two cases: (1) the sky at $ |b| \lesssim 5^{\circ}$ where mean $T_{\textrm{sky}}\sim 600$ K and (2) the sky at $ |b| \gtrsim 5^{\circ} $ , where mean $T_{\textrm{sky}}\sim 270$ K; that is, $T_{\textrm{sys}} = 630$ and 300 K, respectively, as shown in Figure 2.

Intrinsic pulses are broadened due to a variety of effects, as discussed earlier. As detailed in Lorimer & Kramer (Reference Lorimer and Kramer2012), the total smearing time ${\tau _{\textrm{tot}}}$ is the quadratic sum of the finite sampling time ${\tau _{\textrm{samp}}}$ , the residual dispersive smearing due to finite frequency channel ${\tau _{\textrm{chan}}}$ , the dispersive smearing across the full recording band due to finite DM steps in trial DM values ${\tau _{\textrm{dm}}}$ , and the dispersive smearing resulting from piece-wise linear approximation of the quadratic dispersion law in the sub-band dedispersion algorithm employed in searches ${\tau _{\textrm{sub}}}$ . Figure 2 summarises these for our current first-pass processing. The planned second-pass search will significantly enhance the search sensitivity by processing the full observation (4800 s) and the use of more optimal DM steps, i.e. many more trial DMs than that used in current search.

Table 2.

Dedispersion plan for the first-pass SMART processing.

The columns 1 and 2 denote the ranges in dispersion measure, between ${\rm DM} _{\rm min}$ and ${\rm DM} _{\rm max}$ , with a DM step size of $\unicode{x03B4} {\rm DM}$ , resulting in $N _{\rm DM}$ trial DM values. The down sampling factor is denoted by $d_s$ , i.e. the factor which the temporal resolution is averaged to yield a net resolution $\Delta t _{\rm eff}$ .

As evident from the figure, for our current first-pass processing, the total smearing time is dominated by finite DM steps; this sub-optimal choice was made in an effort to maximise the number of observations that can be processed to completion toward a shallow all-sky survey within available computing resources. The dedispersion plan utility used is shown in Table 2. In effect, we progressively downsample the data five times over the DM range searched, each time making the DM step size coarser. At ${\rm DM} \gtrsim 3$ ${\textrm{pc cm}^{-3}}$ , the dispersive smearing time within the 10-kHz channel is larger than the native sampling time (100 $\unicode{x03BC}$ s) but still a smaller contribution to the total smearing time, compared to that due to the DM step size. As a result, the net smearing time ${\tau _{\textrm{tot}}}$ displays a step-wise increase as shown in Figure 2, given our dedispersion plan. At very low DMs $\lesssim$ 10 ${\textrm{pc cm}^{-3}}$ , $\tau_{\textrm{tot}} \sim 0.7$ ms but increases to $\sim$ 10 ms at DM $\sim$ 100 ${\textrm{pc cm}^{-3}}$ . In essence, our first-pass search severely compromises the sensitivity to millisecond pulsars (MSPs) at larger DMs and shorter periods, that is, it is currently sensitive to MSPs at $\text{DM}\lesssim$ 30 ${\textrm{pc cm}^{-3}}$ and $P \gtrsim$ 10 ms. As shown in Figure 2, at those larger DMs, the smearing due to scattering (i.e. pulse broadening) can also be significant. The broadening time here is based on the empirical relation in Bhat et al. (Reference Bhat, Cordes, Camilo, Nice and Lorimer2004), which is mostly relevant for pulsars near the plane. As is well known, these scattering estimates can be uncertain by more than an order of magnitude, denoted by the grey shaded region.

The theoretical sensitivity is shown in Figure 2 for different periods, $P=1.0$ , $0.1$ , $0.01$ and $0.001$ s. In all these calculations, we have assumed a duty cycle of 3%, that is, $W_{\rm eff}/P=0.03$ . For each period, a pair of curves are shown: one for the best-case scenario, that is, searches away from the plane, where $T_{\textrm{sys}}\sim 305$ K; and the second for the sky near the plane where the mean ${T_{\rm{sys}}}$ is twice as high. In either case, the sensitivity is maximum for long-period pulsars, at low to moderate DMs of $\lesssim$ 50 ${\textrm{pc cm}^{-3}}$ , and toward $|b| \gtrsim 5^{\circ}$ .

Figure 3.

Tied-array beam traces through the MWA primary beam for SMART observations. Three example pointing directions for each observation are traced including 1 h before and 1 h after the 80-min observation. The target trace (rotating clockwise as time advances) is coloured pink to represent the trajectory before the observation, red during the observation, and blue after the observation is complete. North is at $0^{\circ}$ and the azimuth angle increases to the East. The colour scales are the same for each subplot, highlighting the sensitivity penalty incurred for observing away from zenith.

With our first-pass processing scheme (i.e. 10-min integrations and a sub-optimal dedispersion plan), we reach a limiting sensitivity of $S_{\textrm{min}}\sim 7\text{--}12\,{\textrm{mJy}}$ for long-period pulsars, and $\sim 12\text{--}25\,{\textrm{mJy}}$ for MSPs at low to moderate DMs. For the proposed deep-pass processing (i.e. $\sim$ 80-min integrations and a more granular dedispersion plan), we can achieve a limiting sensitivity of $S_{\textrm{min}} \sim 2\text{--}3\,{\textrm{mJy}}$ for long-period pulsars and $\sim$ $5\text{--}10\,{\textrm{mJy}}$ for MSPs at low to moderate DMs. In this case, the SMART survey sensitivity is comparable to that of the LOTAAS survey in the northern hemisphere. While LOTAAS can be twice as sensitive as SMART for long-period pulsars, the sensitivity for $P \lesssim$ 10 ms is almost similar, owing to a lower degradation in sensitivity in the SMART band. Compared to the Southern Pulsar Survey of the 1990s at 430 MHz (i.e. a wavelength of 70 cm), also known as the Parkes 70cm survey (Manchester et al. Reference Manchester1996), the SMART survey is $\sim$ 3–5 times more sensitive, especially for pulsars at DMs $\lesssim$ 100 ${\textrm{pc cm}^{-3}}$ and spectral index $\unicode{x03B1} \lesssim -1.5$ . Even the ongoing shallow survey is comparable to the 70 cm survey in theoretical sensitivity, and if at all, slightly more sensitive to steep spectrum pulsars with no turnover down to $\sim$ 100 MHz. This provides a strong motivation to undertake a full-scale pulsar survey with the MWA.

2.5. Effective dwell time and sensitivity

Unlike most other pulsar surveys, where single-dish telescopes are used to track targeted positions for small time intervals (e.g. HTRU, GBNCC), the SMART observations are drift scans, where the primary beam is pointable but static in horizontal coordinates (azimuth and zenith angle) once an observation starts and the sky moves through the FoV. When forming TABs, we track the sky position as it moves through the MWA primary beam and as a consequence not all TABs necessarily remain within a sensitive part of the primary beam for the full 80-min duration.

The amount of time spent within an individual observation FWHM depends both on the observing declination (i.e. where the primary beam is pointed) and the target source position to be tracked with a TAB. As an example, in Figure 3 we plot some representative TAB pointings along with the primary beam response for the same observations as in Figure 1 in horizontal coordinates. As already noted, our sensitivity drops substantially as we observe at larger zenith angles, which we visualise by having the colour scale represent the zenith-normalised primary beam power as a proxy for sensitivity. Secondly, the TAB pointing directions are traced before, during, and after the 80-min observation, which highlights that not all targeted positions remain in a usable part of the primary beam. These effects highlight at least three points for consideration: (1) it will be an inefficient use of resources to track certain pointings for the full observing duration, (2) tracking pointings naively for the full duration, especially if a significant fraction of the time is spent below the 10% power point, may actually reduce sensitivity to pulsars, and (3) the full-sky sensitivity will be patchy regardless of TAB forming strategies (although this is partly mitigated by having observations overlap by $\sim$ 20% at the central frequency). To address (1) we can estimate the time a source remains in a reasonable power range of the primary beam and only form TABs from the appropriate subset of voltages recorded (e.g. while the target source is not in a null). For (2) we must strike a careful balance between achieving maximal sensitivity (by cutting off parts of the TAB) and dwell time (which benefits searches for longer-period pulsars and single pulse events). The consequence of (3) is unavoidable given the telescope configuration and observing strategy employed, but is quantifiable.

We can evaluate the relative sensitivity (assuming a 80-min track for a given TAB) by summing the primary beam response power at discrete time steps, where we use our current best Full Embedded Element (FEE) model (Sokolowski et al. Reference Sokolowski2017), normalised to the equivalently summed power that would represent the best possible dwell time and sensitivity combination. For our purposes we define this quantity as the sum of the primary beam power at zenith for the full observing duration (i.e. imagining we can track an equatorial position with full zenith sensitivity). This is useful as it scales the effective sensitivity to a quantity close to what a single-dish steerable telescope could achieve. In Figure 4 we present these effective sensitivity maps, in equatorial coordinates, for the same example observations used in Figure 3.

Figure 4.

Effective sensitivity maps, assuming a full 80-min tracking and integration for a given TAB sky position. The colour map is normalised to the best possible sensitivity (described in the text), and contours at 25, 50, and 75% are drawn for clarity. Due to the drift scan nature of the observations versus the tracking TABs, we can never achieve the best possible sensitivity. Right Ascension and Declination are marked by the vertical and horizontal curved grid lines, respectively.

3.1. Preprocessing and beamforming

The main step in the preprocessing stage involves processing VCS data so they can be calibrated and coherently combined to produce beamformed time series at the native resolution of 100- $\unicode{x03BC} {\rm{s}}$ /10-kHz of the VCS. The array calibration is performed using one of the standard calibrators (e.g. 3C444), recorded in the visibility mode at the default 0.5-s/40-kHz resolution, where complex gain solutions (amplitude and phase) are obtained for each of the 128 tiles, for every coarse channel (1.28 MHz wide), using the Real Time System (RTS) software package. The procedure is essentially similar to those employed for other VCS observations (e.g. Swainston et al. Reference Swainston2021). The calibration solutions can then be used to coherently combine the voltage data in phase using the tied-array beamformer, the conceptual details and implementation of which are detailed in Ord et al. (Reference Ord2019). The functionality was enhanced, and GPU parallelised, in preparation for SMART data processing (Swainston et al. Reference Swainston2022).

The beamformed data are written as Stokes I at 100- $\unicode{x03BC} {\rm{s}}$ /10-kHz resolutions. The current implementation allows processing 120 coarse-channel beams at once, that is, 5 full-bandwidth (30.72 MHz) beams, resulting in a data rate of 87 GB beam $^{-1}$ for a 10-min observation. For each survey pointing, this amounts to $\sim$ 500 TB in beamformed data. These data are equivalent to that would emerge from the standard pulsar backends and so can be processed using standard pulsar search packages. Typically, data would be processed to generate radio frequency interference (RFI) masks; however the superb radio-quiet environment at the telescope site and preferential observing during the nightly hours (and within an hour of the source transit) make this step not essential for the SMART data. In most cases, data are minimally affected by RFI, and consequently no RFI-related processing is carried out in the ongoing first-pass processing.

The large FoV of the MWA means excellent prospects for detecting multiple known pulsars within each pointing, which is also important for crucial data quality checks and initial assessment of array calibration and tied-array sensitivity. In short, each SMART observation is processed for known pulsars within the primary beam ( $\sim 610\, {\textrm{deg}^{2}}$ ), using a custom pulsar detection pipeline.

3.2. Search pipeline

The current SMART pipeline includes a GPU-based pipeline for front-end processing (beamforming) and a CPU-based pipeline for downstream (search) processing. The search pipeline is based on the Pulsar Exploration and Search ToolkitFootnote c (PRESTO; Ransom Reference Ransom2001, Reference Ransom2011) pulsar search software suite, with the addition of machine-learning (ML) tools adopted from the LOTAAS classifier (Tan et al. Reference Tan2018 a). This was adopted as a first-pass processing strategy, to ensure an end-to-end working pipeline from the data collection and reordering stage (occurring at the observatory site) to array calibration/quality checks (Pawsey) and search processing (OzSTAR). To encapsulate the full-search workflow, we make extensive use of NextflowFootnote d (Di Tommaso et al. Reference Di Tommaso2017) to manage data input, output, processing tasks, and intermediate or final product creation and tracking.

In the near future, as we transition to full sensitivity searches, the search component will be replaced by a GPU-based implementation. Here we present a detailed breakdown of the current SMART search pipeline, where 10-min data (4.8 TB) are processed from each observation.

3.2.2. Candidate folding

Folding of the candidates is performed using the prepfold tool, which creates the associated candidate files and standard diagnostic plots such as those shown in Figure 6. Since our pipeline uses the LOTAAS classifier, the folding analysis is carried out using the identical parameter setup as in the LOFAR search pipeline; that is, 100 pulse phase bins, 256 sub-bands, 120 sub-integrations for $P>10$ ms, whereas 50 pulse phase bins and 40 sub-integrations for $P<10$ ms. With this, the folded candidate information can be classified and processed using the ML classifier that we have adopted from the LOFAR search.

Figure 6.

Examples of standard PRESTO diagnostic plots of original periodic pulsar candidate detections (left panels), and improved detection plots from follow-up processing for confirmation (right panels). Upper panels are the first pulsar discovered from the SMART, PSR J0036–1033, and the lower panels are the second pulsar, PSR J0026–1955. Initial detections are from 10-min observing durations (first-pass processing), while the confirmation ones are from longer durations of the same initial detection observations.

3.3. Pulsar detection pipeline

3.3.2. Prioritisation and scrutiny of candidates

The candidates that survive the initial ML cull are still mostly dominated by noise and RFI detections, with only a small minority being true pulsar detections. Although all of these candidates are intended ultimately to be visually inspected, we have developed a so-called ‘clustering’ algorithm to prioritise which candidates get inspected first, in order to accelerate the detection of sufficiently bright, new pulsars.

Figure 7.

The theoretical array factor (a proxy for sensitivity) of each tied-array beam towards the pulsar B2327 $-$ 20, with the red cross marking the position of the pulsar (left panel) and the beams in which the pulsar was detected (right panel). SMART observation 1226062160 was used for the demonstration.

Figure 8.

The theoretical array factor (proxy for sensitivity) in the vicinity of PSR J0026–1955 for observation 1226062160, assuming a true position (centre of image) derived from GMRT imaging (cf. Paper II for details). Red crosses mark the position of beams in which it was detected, and the blue dot marks the first detection. A single cross may indicate multiple detections with slightly different periods and DMs.

The clustering algorithm leverages the fact that the tied-array beam of the MWA’s compact configuration is relatively complex, with significant grating lobes located in different parts of the primary beam. Because the spacing between tied-array pointings is equal to the FWHM of the main lobe of the tied-array beam, any sufficiently bright pulsar will likely be detected in multiple beams. For instance, Figure 7 shows a map of multiple detections of PSR B2327 $-$ 20 superimposed on the theoretical sensitivity of each tied-array beam towards the pulsar, as predicted by the array factor formalism developed for the MWA by Meyers et al. (Reference Meyers2017). Since noise candidates will not be correlated across different beams, prioritising similar candidates that appear in multiple beams dramatically increases the likelihood that candidates representing true astrophysical signals will be inspected first.Footnote f

Candidates are considered similar if

1. 1. they appear in at least two adjacent beams,

2. 2. they have periods within 0.5% of each other, and

3. 3. they have DMs within 3 ${\textrm{pc cm}^{-3}}$ of each other.

As a demonstration of the usefulness of the clustering algorithm, we show how it would detect PSR J0026–1955, the second pulsar discovery in the SMART survey (McSweeney et al. Reference McSweeney2022). In reality, the clustering algorithm was not implemented until after PSR J0026–1955 was discovered, but it is interesting to note that the first detection (chronologically) of this pulsar was a grating lobe detection (at the time, the candidates were being served up randomly), which motivated the development of the clustering algorithm in the first place.

The final set of detections of PSR J0026–1955 is shown in Figure 8, on a backdrop of the theoretical array factor (a proxy for sensitivity) towards the pulsar assuming that our current best-fit position is correct. In this case, three of the search beams contained the nominal pulsar position in the main lobe, while several others positioned the pulsar in their respective grating lobes. All of the displayed detections meet the second and third clustering criteria (similar periods and DMs). Therefore, any pair of detections in the same or adjacent beams are considered ‘clusters’, and if the clustering algorithm was in use when this observation was processed, this pulsar would have been picked up immediately in multiple clusters.

The clustering algorithm offers no advantage for relatively weak pulsars that would be detected only in a single (boresight) tied-array beam. Therefore, unclustered candidates are not deleted, only deprioritised.

3.4. Data management and web app

The large number of generated candidates, the complex metadata associated with them, and the desire to distribute the tasks of data processing, candidate rating and candidate follow-up, motivated the implementation of a relational database to track the progress of the SMART survey and coordinate processing efforts. The database, implemented in PostgreSQL, is comprised of a set of tables containing metadata for

1. 1. MWA observations (e.g. primary beams, tied-array beams, candidates);

2. 2. software (e.g. for beamforming, searching, ML classification), including versioning information;

3. 3. candidate ratings;

4. 4. pulsars;

5. 5. users; and

6. 6. supercomputer facilities.

The users, along with their database access privileges and authentication, are managed by a subset of tables which interface with website front end implemented in Django. Both the database and the website are hosted by Data Central.Footnote h

Once an observation has been processed and the candidates have been subjected to the first-pass ML cull (Section 3.3.1), both the metadata of the remaining candidates as well as the candidates themselves (i.e. PRESTO.pfd files and the associated diagnostic plots) are uploaded to Data Central. The uploaded candidates are then available for users to rate via the web interface (Section 3.3.3).

As described above, candidates can then be sorted by their average rating, and followed up at will by any authorised user. Before following up a candidate, the user may ‘claim’ it by clicking a button in the candidate list. This feature is designed to prevent multiple users from following up the same candidate and unnecessarily duplicating effort. The decentralised design allows members of the SMART collaboration from different research institutions to work through the SMART data set without the need for someone to oversee and coordinate the different groups’ activities.

4.1. Improved detection

For our ongoing shallow survey, processing the full 80-min observation itself readily provides an avenue for confirmation. If the source is genuine and a steady periodic emitter, this should result in a three-fold improvement in S/N. The improvement will be reduced if it is an intermittent source; for example, a pulsar with large nulling fraction. Both these possibilities are exemplified in Figure 6, which shows the original discovery plots along with the improved detections for PSRs J0036–1033 and J0026–1955. The full 80-min observations (42 TB) containing the original detection can be processed and searched over a restricted range in P and DM using the PRESTO prepfold routine. The observations were also processed using the pdmp routine within PSRCHIVE pulsar data processing suiteFootnote i (Hotan, van Straten, & Manchester Reference Hotan, van Straten and Manchester2004; van Straten, Demorest, & Oslowski Reference van Straten, Demorest and Oslowski2012), to provide a cross-check and a more accurate DM. This is equivalent to undertaking a longer observation for confirmation. For many of our candidates, this readily provides effective ways of confirming or rejecting a candidate, and eliminates the need for securing additional telescope time that most other surveys typically require.

While the long dwell time of 4800 s should in principle result in an increased sensitivity to sporadic or intermittent pulsars, our current first-pass processing does not necessarily benefit from this. Given this, the discovery of PSR J0026–1955 in the first 10 min of observations, a pulsar with long-duration nulls and a nulling fraction of $\sim$ 77%, was remarkably fortuitous (see Figure 6). Details of the discovery, including an analysis of sub-pulse drifting, are reported in McSweeney et al. (Reference McSweeney2022). As mentioned therein, this pulsar turned out to have already been reported as a candidate in the GBNCC survey but was blindly (and independently) discovered in the SMART survey data.

Figure 9.

MWA localisation of PSR J0026–1955 by performing a dense grid around the initial pulsar position from the discovery observation. The source position $\rm (RA, Dec)=(00^h26^m37.5^s, -19^{\circ} 56^{\prime} 24.9^{\prime\prime})$ is $\approx 32^{\prime\prime}$ offset from uGMRT-determined position (cf. Paper II for further details). Observations were made using the extended MWA array (Phase II, with $\sim$ 6 km maximum baseline). The uncertainties in the MWA position is $\sim$ $12^{\prime\prime}$ (i.e. about one-tenth of the tied-array beam size, shown as dashed circles on the left panel).

4.2. Improved positional determination

As outlined in Section 2.3, the tied-array beam size for SMART is $\sim$ $23^{\prime}$ . Therefore a more accurate position is essential both for improved detection (i.e. re-beamforming on a more exact sky position) and to facilitate effective follow-ups with other (and more sensitive) telescopes, particularly at higher frequencies where the beams are narrower, even with single-dish telescopes such as Parkes. This would typically involve making multiple re-observations to form a grid around the nominal candidate position. The SMART survey design where the sky is densely sampled (at a rate comparable to, or slightly better than, the Nyquist; Figure 1), allows this to be achieved via reprocessing of the original survey observation, where a dense grid of pointings encompassing the initial position is used for improved positional determination. An example is shown in Figure 9 for the case of PSR J0026–1955. In general, for an initial detection with a modest significance of $\rm S/N \sim 10$ , we may expect a positional accuracy $\sim$ 1-2 $^{\prime}$ through this exercise. In practice, archival VCS data, if available, can also be suitably exploited to progressively further improve the position. In an ideal scenario, where data recorded from all three different configurations are available, an improvement of the order of nearly two orders of magnitude can be achieved through this procedure, as demonstrated in Swainston et al. (Reference Swainston2021).

4.3. Polarimetry

The VCS recording allows the reprocessing of discovery observations to generate full polarimetric beamformed time series, which can be analysed using standard pulsar packages such as DSPSRFootnote j (van Straten & Bailes Reference van Straten and Bailes2011) and PSRCHIVE, for full Stokes profiles. These beamformed MWA data were obtained using the procedures described in Ord et al. (Reference Ord2019) and Xue et al. (Reference Xue2019). The Faraday rotation measure synthesis technique (Brentjens & de Bruyn Reference Brentjens and de Bruyn2005) can then be applied to estimate the rotation measure (RM).

As an example, Figure 10 shows polarisation data for pulsar J0026–1955, obtained by reprocessing the original discovery observation. This yielded an RM estimate of $ 3.65 \pm 0.09 $ ${\textrm{rad m}^{-2}}$ . After correcting for Faraday rotation, linear and circular polarisation was detected. The pulsar exhibits significant amount of linear polarisation but only a small amount of circular polarisation. We attempted to fit the rotating vector model (Radhakrishnan & Cooke Reference Radhakrishnan and Cooke1969) to the position angle (PA) of the linear polarisation across the on-pulse window, in order to constrain the viewing geometry, $(\unicode{x03B1}, \unicode{x03B2})$ , where $\unicode{x03B1}$ is the angle between the magnetic and rotation axes, and $\unicode{x03B2}$ is the impact angle of the magnetic axis on the line of sight. In the absence of relativistic effects, the PA curve is expected to be steepest in the centre of the pulse profile, with slope $d\unicode{x03C8}/d\unicode{x03D5} = \sin\unicode{x03B1} / \sin\unicode{x03B2} \approx 2.4$ , where $\unicode{x03C8}$ is the PA at phase $\unicode{x03D5}$ .

Figure 10.

Polarimetric profiles of PSR J0026–1955 obtained by reprocessing the discovery observation at 155 MHz. The black, red, and blue curves in the lower panels show the total intensity, linear, and circular polarisation, respectively. An RM estimate of $ 3.65 \pm 0.09\,{\textrm{rad m}^{-2}}$ was obtained, and the data were corrected for Faraday rotation.

5.1. Long-period pulsars

The discovery of two new pulsars from the processing of a small fraction of survey data hints at the potential for many new pulsar discoveries from a deeper survey that will take advantage of the full 80-min observation. To estimate the survey yield, we have performed survey simulations, using the formalism outlined in Xue et al. (Reference Xue2017). The analysis made use of the popular simulation package PsrPopPy (Bates et al. Reference Bates, Lorimer, Rane and Swiggum2014) that was developed from the original pulsar simulation software PSRPOP by Lorimer et al. (Reference Lorimer2006). The simulations take into account the sky dependence of the system temperatures at low frequencies ( $T_{sky} \propto \nu^{-2.55}$ ), as well as the loss in the array gain (G) expected at large zenith angles, modelled as $G(\unicode{x03B8} _z) = G_{\rm max} {\rm cos}(\unicode{x03B8} _z)$ , where $\unicode{x03B8} _z$ is the zenith angle and $G_{\rm max}$ is the gain at $\unicode{x03B8}_z=0$ . We simulated a population of $1.6\times10^{5}$ Galactic canonical pulsars, extrapolated from Parkes Multibeam Pulsar Survey (Manchester et al. Reference Manchester2001) detections. The luminosity distribution of the canonical pulsar population follows a log-normal distribution $\langle{\rm log}_{10}L\rangle=-1.1$ , $\sigma[{\rm log}_{10}L]=0.9$ , where L is the radio psuedo-luminosity in units of $\textrm{mJy kpc}^{2}$ (Faucher-Giguère & Kaspi Reference Faucher-Giguère and Kaspi2006). The Galactic radial density distribution follows the Yusifov & Küçük (Reference Yusifov and Küçük2004) model. With the caveat that our understanding of the pulsar luminosity function and beaming fraction is limited, we project the deep survey to reach a limiting sensitivity of $\sim$ 2–3 ${\textrm{mJy}}$ , with a potential net yield of $310 \pm 100$ new pulsar discoveries (see Figure 11). This projection mainly applies to the population of long-period pulsars and does not account for other classes of pulsars such as sporadic emitters (e.g. RRATs), or millisecond and binary pulsars, whose populations are hard to model or simulate.

Assuming an isotropic distribution of our simulated local pulsar population ( $\rm DM \lesssim 250$ ${\textrm{pc cm}^{-3}}$ ), and scaling for the current (first-pass) search sensitivity (i.e. one-third of the deep-pass sensitivity), and the fraction of data for which the candidate scrutiny has been completed ( $\sim$ 5%), we may expect $\sim$ 3–5 pulsars. The detection rate at this early stage of SMART thus appears to be in line with this general expectation. While this may seem fortuitous, the unique advantages of the SMART pulsar survey, especially the accessibility to the southern hemisphere, the radio-quiet environment, and the survey parameters (e.g. long dwell times and high time/frequency resolutions), offer excellent prospects for new pulsar discoveries, provided the substantial processing challenges can be addressed.

5.2. Millisecond pulsars

Even though the detection sensitivity to MSPs is significantly reduced in our current shallow pass of the survey (owing to the use of coarse or sub-optimal DM step sizes; see Figure 2), the second-pass processing, where we plan to employ more optimal DM searches with a finer step size in DMs, is expected to yield a substantial improvement in sensitivity, particularly at low to moderate DMs, out to $\lesssim$ 50 ${\textrm{pc cm}^{-3}}$ . At DMs $\gtrsim$ 70 ${\textrm{pc cm}^{-3}}$ , and especially in regions near the Galactic plane and toward the centre, scatter broadening is expected to result in sensitivity degradation, given the strong frequency dependence (pulse broadening time, $\tau _d \propto \nu^{-3.9}$ ; cf. Bhat et al. Reference Bhat, Cordes, Camilo, Nice and Lorimer2004), due to which $\tau _d \gtrsim $ 10 ms, which, for millisecond pulsars, can be a substantial fraction of the rotation period. Using PsrPopPy, we simulated a population of $3\times10^{4}$ MSPs with P and DM distributions essentially derived from the HTRU intermediate latitude pulsar survey (Levin et al. Reference Levin2013), and with a luminosity limit of $L_{1400} \sim 0.2\,{\textrm{mJy kpc}^{2}}$ . This corresponds to a limiting flux density $\sim$ 10 ${\textrm{mJy}}$ at 150 MHz, assuming a spectral index of $\unicode{x03B1} = -1.8$ (and a distance of $\sim$ 1 kpc), and thus in principle detectable provided there is no significant degradation from dispersive smearing or temporal broadening from scattering.

Figure 12.

Simulated pulsars detectable in an all-sky pulsar search with the MWA’s 140–170 MHz band with a dwell time of 4800 s. The shaded region represents the MWA’s visible sky, that is, the sky south of $+30^{\circ}$ in declination. The black filled circles denote the long-period pulsars, whereas millisecond pulsars detectable in high-sensitivity searches (e.g. using the CDMT) are shown as colour filled circles. The colour scale indicates DM in units of ${\textrm{pc cm}^{-3}}$ .

As with the population of long-period pulsars, this analysis accounted for the sky dependence of ${T_{\textrm{sky}}}$ , non-uniformity in the array gain, and strong frequency scaling of scattering, which is especially important for MSPs. For example, using some preliminary dedispersion plan estimates for the second round of processing (i.e. the deep search), where we assume a typical plan would involve DM steps of 0.01 ${\textrm{pc cm}^{-3}}$ up to 54 ${\textrm{pc cm}^{-3}}$ and 0.02 ${\textrm{pc cm}^{-3}}$ out to 107 ${\textrm{pc cm}^{-3}}$ , our simulations predict 55 detectable MSPs above our detection threshold, and hence $\sim$ 15 new MSP discoveries. However, a substantial increase is forecast in simulations that closely emulate the higher sensitivity attainable through more optimal searches that make use of coherent dispersion measure trials (CDMT), which is equivalent to the use of finer DM steps of 0.002 ${\textrm{pc cm}^{-3}}$ , and will limit residual DM smearing to $\sim$ $150\,{\unicode{x03BC}{\rm s}}$ (comparable to $\sim$ $100\,{\unicode{x03BC}{\rm s}}$ native resolution of the VCS). In essence, this means that full-scale, high-sensitivity searches employing the implementation of CDMT, if feasible for SMART, can potentially lead to the discovery of as many as $\sim$ 30 MSPs.

The simulated population of $\sim$ 70 MSPs, along with the simulated population of long-period pulsars (see Section 5.1), is shown in Figure 12. Our simulation analysis did not consider a large population of MSPs discovered in recent (and highly successful) Fermi-directed targeted searches (Deneva et al. Reference Deneva2021, and references therein). Even so, the detectable population of MSPs is almost twice the currently known population within $\rm DM \lesssim 100$ ${\textrm{pc cm}^{-3}}$ , which means a net MSP yield that is competitive to that from the highly successful Parkes HTRU survey. Indeed, as evident from Figure 12, the detectable population of MSPs is limited to $\rm DM \lesssim 70$ ${\textrm{pc cm}^{-3}}$ , which is reconcilable given the expected pulse broadening times of $\tau _d \gtrsim $ 10 ms toward such moderate DM pulsars at the low frequencies of the MWA (e.g. Kirsten et al. Reference Kirsten2019). Consequently, the vast majority of MSPs discovered will likely be suitable for high-precision timing applications such as pulsar timing arrays.

6.1. Beamforming and sensitivity optimisation

As discussed earlier in Section 2.5, the tied-array beamforming strategy warrants some more careful thought in order to maximise sensitivity while also reducing needless processing. Inevitably, this produces an uneven sensitivity threshold across the sky due to both primary beam pointing effects and effective dwell time. These considerations are also important when estimating survey-wide statistics. We are formulating a more efficient beamforming scheme that takes into account these technical details, which will be presented in a subsequent paper detailing the second-pass survey processing.

6.2. Dedispersion planning and RFI mitigation strategies

For the first-pass survey processing described in this paper, the dedispersion plan outlined in Table 2 is adequate for all observations. In contrast, a slightly more sophisticated plan may be required for the second-pass processing to accommodate the eight-fold increase in observation length and to provide increased sensitivity to shorter-period pulsars. We are actively developing a sensible strategy that balances our sensitivity goals and the relatively large computational costs associated with dedispersing MWA VCS data, especially since we would essentially be producing $\sim$ 10 $\times$ as many DM trials.

In addition to revisiting the dedispersion plan, we will also incorporate a more careful approach to excising or mitigating RFI (both periodic and impulsive). The observatory site is exceptionally RFI-quiet (owing to the geographical location and radio-quiet zone status), hence the first-pass processing did not include any active RFI mitigation other than what is naturally gained by forming TABs (where off-axis RFI is ‘phased out’). We are currently examining the periodic RFI environment by processing observations taken throughout the SMART observing semesters and using a standard PRESTO-based approach to find bright, common terrestrial signals by searching for periodic ‘candidates’ in the zero-DM topocentric time series data. Once we collect this information, we will apply the masks (after appropriate barycentric corrections are made) during the periodicity search pipeline. Additionally, there can occasionally be bursts of narrowband interference (e.g. air-craft and satellites in TAB grating lobes) that could severely affect our data quality for short periods of time. There are several software preprocessing solutions to this kind of RFI (e.g. Eatough, Keane, & Lyne Reference Eatough, Keane and Lyne2009; Men et al. Reference Men2019; Morello, Rajwade, & Stappers Reference Morello, Rajwade and Stappers2022), which we will explore in parallel to the periodic RFI mitigation strategies. Empirically, VCS data are remarkably clear of impulsive/narrowband RFI in the SMART observing band, and data excision is $\ll$ 10% for a typical observation.

6.3. Searches for long-period pulsars and sporadic emitters

The long dwell times of SMART make it particularly amenable to the application of fast-folding algorithms that offer significantly higher sensitivity to pulsars with rotation periods $\gtrsim$ 10 s (e.g. Morello et al. Reference Morello2020). Such slow-spinning pulsars are likely to be near the radio emission ‘death lines’ so can be invaluable in gaining useful insights into the intricacies of the pulsar radio emission process. Recent applications of this algorithm in Parkes and Arecibo searches have led to the discoveries of pulsars with $P >$ 10 s (Morello et al. Reference Morello2020), or very weak pulsars ( $S_{1400}\sim10$ $\unicode{x03BC}$ Jy) with a $\sim$ 2% duty cycle (Parent et al. Reference Parent2018). These, and other recent discoveries such as a 76-s pulsar with MeerKAT (Caleb et al. Reference Caleb2022), provide a strong motivation for undertaking fast-folding searches. The low levels of RFI at the observatory site are particularly advantageous for this.

The SMART survey dwell time is substantially longer than those of previous-generation southern-sky surveys, particularly at high- $|b|$ parts of the sky, where it is 20 times longer than the HTRU survey (Keith et al. Reference Keith2010) and 40 times longer than the southern pulsar survey (Manchester et al. Reference Manchester1996). It is also 40 times longer than that of the ongoing GBNCC survey (Stovall et al. Reference Stovall2014) that covers the sky north of $-55^{\circ}$ in declination (Table 1). Considering this, detection prospects are promising, especially given the $\sim$ 2–3 ${\textrm{mJy}}$ limiting sensitivity that the SMART can attain for long-period pulsars (Section 2.4) and negligible degradation in signal strength due to dispersion and pulse broadening effects.

As described earlier, the long dwell times also increase the search sensitivity to objects that emit sporadically, such as RRATs and giant-pulse emitters (e.g. the Crab pulsar), which can be more effectively detected by searching for individual dispersed pulses, and will be part of the second-pass processing.

6.4. Searches for binary and millisecond pulsars

The long dwell times, and high time and frequency resolutions, of the SMART can also be exploited, in principle, to search for binary and millisecond pulsars. However, a full-scale acceleration search can be prohibitively expensive at the low frequencies, given the very large number of DM and acceleration trials that are required (e.g. typically $\sim$ $10^4$ up to 250 ${\textrm{pc cm}^{-3}}$ , and $\sim$ $2400$ across $\pm 100\,{\textrm{m s}^{-2}}$ ). Compared to the HTRU-south low-latitude survey, which has been successful in finding such systems (e.g. Cameron et al. Reference Cameron2020), the cost of searching SMART data can be more than an order of magnitude greater. The successful detection of several MSPs and the double pulsar in our initial census (cf. Paper II) makes such searches worthwhile.

An inherent limitation in the searches for such short-period pulsars is the significant degradation in sensitivity due to substantial dispersion smearing (relative to rotation periods) despite our 10-kHz channels. Fortunately, this can be alleviated by using CDMT-based searches (Bassa et al. Reference Bassa2017). Recording in 24 $\times$ 1.28-MHz channels makes the SMART data highly amenable to the application of CDMT searches, and can result in a substantial increase in detection sensitivity to short-period millisecond pulsars. Integration of this novel method, and benchmarking on prospective HPC clusters with significant computational resources (e.g. Pawsey’s emerging Setonix cluster) is also part of our future processing plans, although a full-scale processing may have to await access to sufficient computational resources. We are also exploring publicly available, GPU-enabled Fourier domain acceleration search software (e.g. AstroAccelerate; Armour et al. Reference Armour2020) as a drop-in replacement for PRESTO’s CPU-based accelsearch.

Regardless, the high cost of such computationally-intensive searches will likely necessitate a multi-pass processing strategy; for instance, an initial pass involving acceleration searches, but limited to a modest number of acceleration trials (e.g. $\sim$ 150 to cover $\pm 6\,{\textrm{m s}^{-2}}$ ), thereby retaining sensitivity to short-period objects ( $P \lesssim 10$ ms) but with the binary orbital period, $P_b \gtrsim 5$ d (i.e. with low-mass white dwarf type companions). Full-scale acceleration searches that target binary systems such as PSR J0737 $-$ 3039 or PSR J1757 $-$ 1854 with $P_b \lesssim 5 $ h (i.e. requiring $\sim$ 2400 trials spanning across $\pm 100\,{\textrm{m s}^{-2}}$ ) are hence deferred to the longer-term future. Such searches will be primarily limited to the regions around the Galactic plane, at least initially, thus processing only a fraction of the SMART data (e.g. sky within $|b|\lesssim5^{\circ}$ ). Such a multi-pass strategy is also motivated by the demonstrated success of HTRU-south, which has led to the discovery of exotic systems such as PSR J1757–1854 (Cameron et al. Reference Cameron2018) and wide-orbit double neutron-star system (Sengar et al. Reference Sengar2022). In any case, notwithstanding the high computational cost, the high-profile scientific applications of such rare systems make similar full-scale acceleration searches scientifically compelling for the SMART data. The long-term scientific dividends of such systems are vividly demonstrated by Kramer et al. (Reference Kramer2021) through the 16-yr timing analysis of the double pulsar, enabling the most stringent tests of general relativity and alternative theories of gravity.