AI pipelines must resolve SKA's terabit-per-second data bottlenecks to enable real-time discovery

The paper reviews how convolutional networks perform automated source detection and RFI flagging on simulated SKA-Mid data while generative models accelerate calibration and imaging. It further proposes reinforcement learning for dynamic scheduling and federated learning to handle distributed processing across the SKA's global sites. These approaches directly target the shift from post-observation analysis to real-time decision making at terabit rates. Mainstream coverage has overlooked the specific failure modes that arise when models trained on current arrays encounter SKA's higher dynamic range and ionospheric variability. Existing SKA precursor pipelines at MeerKAT already show 15-20% completeness drops at faint flux levels when RFI environments change, a gap the preprint acknowledges but does not quantify for full SKAO scale. Physics-informed neural networks and uncertainty-aware architectures are highlighted as necessary constraints yet remain at the conceptual stage without end-to-end benchmarks on SKA-scale volumes. Next steps require closed-loop tests on precursor arrays with injected anomalies to measure discovery latency and false-positive rates before 2029 construction milestones.