Probabilistic graph neural networks aim to monitor aquaculture ponds under real-time policy limits

A developer built a sensor network across several fishponds in a cooperative near their city to track water quality. They found that per-sensor forecasting models missed cross-pond effects, since shared water channels let problems such as ammonia spikes spread between ponds. Treating the sensor layout as a graph, with ponds as nodes and hydrological connections as weighted edges, allowed the model to capture those dependencies. The work extends this to probabilistic inference and to enforcing regulatory constraints such as discharge limits and antibiotic restrictions in near-real time. The article describes the resulting architecture and lessons from testing it.