Scalable 1D/2D Stormwater and Surface‑Flood ML Modelling with HPC.

Urban areas face increasing flood and pollution risk due to climate change, urbanisation and ageing infrastructure. Traditional hydraulic simulators (Saint‑Venant/Navier–Stokes solvers) are too computationally intensive to support collaborative real‑time planning. Building on previous 1D stormwater ML work, this project will develop and train graph‑neural‑network models that emulate coupled 1D/2D hydraulic simulations. Using synthetically generated networks with real digital elevation models and distributed rainfall, the researchers will teach the model to predict both network flows and surface flooding, including interactions between them. The training dataset (~5 TB) will be ~100× larger than in the development project, requiring high‑throughput, multi‑GPU training on LUMI‑G. The researchers will port the InflowGo NN framework to support 250 concurrent MI250X GPUs, compile torch_sparse, and conduct ~100 experiments (1–10 days each) to explore model architectures and hyper‑parameters. The resulting models will enable collaborative planning, ensemble forecasting and digital‑twin services for utilities and municipalities, dramatically improving resilience and public safety.