SKA-Representative Datasets

• SKA-representative datasets are rigorously constructed products that replicate the data complexity, volume, and instrumental characteristics of the SKA.
• They incorporate realistic sky models, calibration products, and noise characteristics to benchmark imaging, RFI mitigation, and source-finding algorithms.
• Their applications span continuum imaging, HI spectral line studies, time-domain analytics, and co-design benchmarks to ensure methodological reproducibility.

An SKA-representative dataset is a rigorously constructed data product—simulated or observationally acquired—that models the data complexity, volume, and instrumental characteristics of the Square Kilometre Array (SKA) for the purpose of developing, benchmarking, or validating algorithms and pipelines destined for SKA-scale science operations. Such datasets are foundational to pipeline calibration, RFI mitigation, source-finding, and reproducibility frameworks in the era of large-scale radio astronomy, enabling stress testing of software, verification of scientific fidelity, and planning for data transport and storage requirements under realistic constraints.

1. Key Dataset Types and Scientific Motivations

SKA-representative datasets span the major use cases of the SKA, including continuum imaging, HI spectral line (galaxy and intensity mapping), weak lensing shape measurement, time-domain astronomy, heliophysics, and sustainable data processing benchmarks.

Categories of SKA-representative datasets

Domain	Example Dataset (Paper)	Scope/Format
Continuum Imaging	SDC1 (Bonaldi et al., 2020, Bonaldi et al., 2018), astroCAMP (Constantinescu et al., 15 Dec 2025)	FITS, OSKAR/MS, HDF5
HI Spectral Line	SDC2 (Hartley et al., 2023), Karabo (Sharma et al., 1 Apr 2025)	FITS cubes, HDF5
Time-Domain/RFI	EDA2 prototype (Grigg et al., 2023)	MIRIAD, FITS, JSON
Weak Lensing/Shape	RadioLensfit SKA1-1000 (Rivi et al., 2022)	MS, visibilities
Co-design/Benchmarking	astroCAMP (Constantinescu et al., 15 Dec 2025)	MS, HDF5, FITS

The primary scientific motivations are: (1) stress-testing calibration and imaging pipelines; (2) validating source-finding and classification algorithms under ultra-high source density and SNR regimes; (3) benchmarking software/hardware for scalability, energy efficiency, and accuracy; (4) enabling comparative studies across teams under controlled, fully specified data and "truth" conditions.

2. Dataset Construction: Scope, Content, and Formats

SKA-representative datasets model or emulate SKA signal chains, incorporating realistic sky models, instrument response, noise, and (sometimes) systematics for algorithmic reproducibility. Dataset features include:

• Full precision visibilities (e.g., OSKAR, MIRIAD, or CASA MeasurementSet) and/or image cubes (FITS, HDF5).
• Rich metadata (antenna geometry, beam models, calibration tables, experiment parameters).
• Reference "truth" catalogues with well-defined source properties for scoring and benchmarking.
• Calibration products (gain/bandpass/delay solutions).
• Inclusion of both time-domain and spectral fidelity (e.g., subsecond time resolution, hundreds of frequency channels).
• Ground-truth RFI or satellite emission features for RFI mitigation testing (Grigg et al., 2023).
• Fidelity and error metrics (e.g., RMSE, PSNR, SSIM, astrometric error) (Constantinescu et al., 15 Dec 2025).

File and directory layouts are standardized, supporting automated workflows and parallel, out-of-core processing.

3. Benchmark Challenges and Representative Data Releases

The SKA Science Data Challenges (SDC) and algorithmic benchmarks define the field standard for SKA-representative datasets.

SKA Science Data Challenge 1 (SDC1) (Bonaldi et al., 2020, Bonaldi et al., 2018)

• Simulated continuum datasets at 560 MHz, 1.4 GHz, 9.2 GHz; 8, 100, 1000 h integrations.
• 32768×32768 FITS images, realistic morphologies from T-RECS, full primary beam application.
• Performance benchmarking based on source detection, flux/position accuracy, completeness/reliability, population classification.

SDC2 (Hartley et al., 2023)

• Full-resolution HI spectral line cube (2000 h, 20 deg², 913 GB), 233,245 galaxies.
• Detailed source model, evolving mass function, spatial clustering, full PSF and noise propagation.
• Development and evaluation cubes with public truth catalogs and standard scoring.

astroCAMP (Constantinescu et al., 15 Dec 2025)

• Suite of 16 OSKAR-simulated SKA-Low visibility datasets (B=130,816 baselines, 1–256 time and 1–256 frequency samples).
• Published as both native OSKAR HDF5 and CASA MS.
• Public reference images (clean/dirty/residual, WSClean/IDG), with quantitative fidelity, energy, carbon, and cost metrics.
• A benchmark for sustainable and efficient hardware/software pipeline co-design.

4. Calibration, Imaging, RFI Mitigation, and Quality Assurance

SKA-representative datasets are calibrated and imaged using best-practice, fully specified workflows:

• External calibrator-based gain solutions (per-antenna complex gains, bandpass/delay solutions), with explicit formulae for visibility and flux conversion (Grigg et al., 2023).
• Time-differencing and trajectory-correlation for moving RFI (satellite) identification, leveraging high-cadence imaging (Δt ~ 2 s) and differencing over 32 s lags to maximize moving-source SNR (Grigg et al., 2023).
• Standardized imaging pipelines (e.g., WSClean, IDG GPU gridder, RASCIL, MIRIAD), with both natural and uniform visweighting.
• Provenance tracking of all workflow stages, with reproducible Snakefile/Python scripts (Constantinescu et al., 15 Dec 2025).
• Strict tolerances for science data quality: PSNR > 40 dB, SSIM > 0.98, dynamic range > 10⁴, astrometric error < 1 arcsec.

Quality assurance is implemented by cross-checks with released "truth" catalogs and by reporting comprehensive summary metrics.

5. Limitations, Approximations, and Domain-Specific Representation

Even the most advanced SKA-representative datasets embed simplifying assumptions:

• Most releases omit calibration artefacts (phase/gain errors, imperfect bandpass, beam rotation), atmospheric/ionospheric fluctuations, and direction-dependent errors for tractability (Bonaldi et al., 2020, Bonaldi et al., 2018, Constantinescu et al., 15 Dec 2025).
• Full spectral bandwidth and sub-second time sampling are typically reduced for storage efficiency, requiring out-of-core analysis for SKA-scale simulation (Constantinescu et al., 15 Dec 2025, Sharma et al., 1 Apr 2025).
• Real instrument effects—such as RFI non-Gaussianity, spatial/spectral leakage, CLEAN artefacts, and non-isoplanatic beams—are only partially included, yielding data that are “ideal” apart from thermal noise or injected RFI (Grigg et al., 2023).
• Sky models are anchored to external simulations (T-RECS, GLEAM, MIGHTEE, P-Millennium, PINOCCHIO), with population properties (size, flux, spectral index) matching precursor/pathfinder observations (Sharma et al., 1 Apr 2025).

A plausible implication is that algorithmic performance on these datasets may overestimate real-world reliability, especially for imperfections absent from the synthetic products.

6. Applications Across Pipeline Validation, Algorithm Development, and Co-Design

SKA-representative datasets are foundational for:

• Pipeline development: Gain/selfcal solvers, beam modeling, RFI flagging, source extraction, morphological classification, and photometry validation, using both standard and machine learning approaches (Constantinescu et al., 15 Dec 2025, Hartley et al., 2023).
• High-performance computing: End-to-end parallelization benchmarking (CPU, GPU, out-of-core), time/energy/carbon accounting, workflow reproducibility, and architecture-software co-design (Constantinescu et al., 15 Dec 2025).
• Methodological reproducibility: Reference outputs and scored truth catalogs underpin cross-team algorithmic comparisons, challenge leaderboards, and open science best practices (Hartley et al., 2023, Constantinescu et al., 15 Dec 2025).
• EoR, cosmology, and Galaxy studies: HI intensity mapping cubes with realistic astrophysical foregrounds and cosmological signal parameterizations (Sharma et al., 1 Apr 2025).
• RFI/environmental analysis: Capturing, flagging, and modeling intended/unintended RFI (e.g., Starlink satellites) for compatibility with SKA-Low science requirements (Grigg et al., 2023).

7. Data Access, Availability, and Future Directions

SKA-representative datasets are publicly disseminated, with full documentation, open-source scripts, and supporting tools:

• SKA Science Data Challenges—datasets, truth catalogues, and scoring tools: https://astronomers.skatelescope.org/ska-science-data-challenge-1/ (Bonaldi et al., 2020, Bonaldi et al., 2018, Hartley et al., 2023)
• astroCAMP reproducibility kit and dataset suite: https://github.com/SEAMS-Project/astroCAMP (Constantinescu et al., 15 Dec 2025)
• Karabo framework for simulation and sky-model generation, supporting direct integration with OSKAR, RASCIL, and other radio astronomy packages (Sharma et al., 1 Apr 2025).
• Continual evolution towards PB-scale simulation and open integration with SKA Science Regional Centres (SRCs) and heterogeneous compute backends.

As SKA-limited science and pipeline requirements evolve, dataset complexity and representativeness are expected to increase—with greater inclusion of calibration/atmospheric errors, direction-dependent gains, fine-grained spectral/time sampling, multiplexed data streams, hybrid sky models, and detailed RFI realizations. Rigorous definition of accepted fidelity and sustainability metrics will guide community co-design and ensure continued methodological reproducibility at SKA scale.