The Square Kilometre Array Observatory (SKAO) will explore the radio sky to new depths in order to conduct transformational science. SKAO data products made available to astronomers will be correspondingly large and complex, requiring the application of advanced analysis techniques to extract key science findings. To this end, SKAO is conducting a series of Science Data Challenges, each designed to familiarise the scientific community with SKAO data and to drive the development of new analysis techniques. We present the results from Science Data Challenge 2 (SDC2), which invited participants to find and characterise 233245 neutral hydrogen (H i) sources in a simulated data product representing a 2000 h SKA-Mid spectral line observation from redshifts 0.25 to 0.5. Through the generous support of eight international supercomputing facilities, participants were able to undertake the Challenge using dedicated computational resources. Alongside the main challenge, ‘reproducibility awards’ were made in recognition of those pipelines which demonstrated Open Science best practice. The Challenge saw over 100 participants develop a range of new and existing techniques, with results that highlight the strengths of multidisciplinary and collaborative effort. The winning strategy – which combined predictions from two independent machine learning techniques to yield a 20 per cent improvement in overall performance – underscores one of the main Challenge outcomes: that of method complementarity. It is likely that the combination of methods in a so-called ensemble approach will be key to exploiting very large astronomical datasets.

Context. The future deployment of the Square Kilometer Array (SKA) will lead to a massive influx of astronomical data and the automatic detection and characterization of sources will therefore prove crucial in utilizing its full potential. Aims. We examine how existing astronomical knowledge and tools can be utilized in a machine learning-based pipeline to find 3D spectral line sources. Methods. We present a source-finding pipeline designed to detect 21-cm emission from galaxies that provides the second-best submission of SKA Science Data Challenge 2. The first pipeline step was galaxy segmentation, which consisted of a convolutional neural network (CNN) that took an H i cube as input and output a binary mask to separate galaxy and background voxels. The CNN was trained to output a target mask algorithmically constructed from the underlying source catalog of the simulation. For each source in the catalog, its listed properties were used to mask the voxels in its neighborhood that capture plausible signal distributions of the galaxy. To make the training more efficient, regions containing galaxies were oversampled compared to the background regions. In the subsequent source characterization step, the final source catalog was generated by the merging and dilation modules of the existing source-finding software SoFiA, and some complementary calculations, with the CNN-generated mask as input. To cope with the large size of H i cubes while also allowing for deployment on various computational resources, the pipeline was implemented with flexible and configurable memory usage. Results. We show that once the segmentation CNN has been trained, the performance can be fine-tuned by adjusting the parameters involved in producing the catalog from the mask. Using different sets of parameter values offers a trade-off between completeness and reliability.