Geometric Deep Learning for Modeling, Prediction and Forecasting in Urban Water Systems

Urban water systems are facing growing pressure from climate change, increasing demographics, and geopolitical tensions. Decision-makers and operators need a novel set of tools to design and execute emergency strategies, develop optimal operation schemes for sustainable water management and plan infrastructural adaptation measures under deep uncertainty. Buoyed by the success of pilot applications, water managers have identified digital innovation as the key factor for a resilient water future. IoT devices and remote sensing are bringing "big data" into the water sector, thus enabling the development of game-changing AI solutions. However, in order to accurately simulate physical processes and predict flood hydrodynamics, AI models must account for the underlying graph structure of water networks and the complex topological properties of the urban environment.

In the past five years, the scientific community has increased efforts to devise Geometric Deep Learning (GDL) solutions for data residing in non-Euclidean domains, such as data over graphs. The most successful class of GDL models are Graph Neural Networks (GNNs), direct extensions of Deep Learning methods to graph data. The promising results of GNNs in several fields demonstrate the benefits of incorporating the underlying network structure of the problem at hand in the data-driven model. Yet, GNN research is still in its infancy and urban water applications require fundamental advances to capture time-varying feature information, grant robustness to topological changes and learn from information over higher level simplicial complexes (e.g. flood cells).

This talk will present the first steps of AidroLab —TU Delft's newly founded lab for AI research in sustainable water management— in developing GDL-based data-driven solutions for the urban water cycle. The main goal of AidroLab is to build fast and accurate data-driven simulation engines for prediction of water network behaviors and forecasting urban flood hydrodynamics. These engines will assimilate data from a variety of sensors and serve as the backbone for end-to-end modular AI solutions that can also process external inputs. The talk will also illustrate the core ideas behind GDL and GNN and discuss possible applications in the broader fields of water resources and hydrologic forecasting.