Radio Interferometry is undergoing an epoch-defining expansion, with many next-generation instruments in final planning or under commissioning (e.g., SKA, ngVLA, ngEHT and their pathfinders). Nevertheless, with these great opportunities come some great challenges, and perhaps most pressingly, we must update our computational approaches to match the updated infrastructure.
Machine Learning approaches are ideal for addressing these ‘big data’ questions, with many new applications being discovered almost daily. The opportunities are practically infinite. One particularly promising approach is the use of Graphical Neural Networks (GNNs).
The traditional approach for imaging radio-interferometric data has been to convert the 3D temporally sampled data to a 2D regular grid, then Fourier transform and iteratively correct for the instrumental effects. However, this approach’s multitude of approximations limits its accuracy and scalability. GNNs provide a powerful alternative by directly operating on the irregular data domains sampled by the interferometer.
GNNs extend neural networks to process data represented as graphs, capturing node features and graph topology. For interferometers, the visibilities can be described as node features on a graph defined by the antenna locations and baseline connections. This provides a morphological match between the data domain and the machine learning framework, massively enhancing convergence compared to operating on gridded data.
Using GNN architectures like Message Passing Neural Networks, we can learn complex non-linear mappings from the graph-structured visibilities to the desired sky image reconstruction, while automatically accounting for direction-dependent effects. Attention and normalization layers allow scaling to extremely large graphs.
We would apply GNN imaging to real data from the SKA pathfinders (MWA, ASKAP) and early science (or simulated) SKA datasets, testing the limits of current computation capabilities. We expect this to become a major focus for the SKA Data Processing pipeline, particularly for scales beyond AA2 in 2026.
The outcome of the PhD would be an innovative new approach to robust, scalable imaging in the SKA era, enabling crucial science applications for SKA, ngVLA, and ngEHT. This experience would provide valuable expertise in cutting-edge GNN development with prospects for broader academic and industry applications.