Digital Twins for Coastal Flooding

Digital Twins (DTs) represent a transformative approach to coastal flood risk management, yet their full realization in coastal hydrodynamics remains challenging due to computational bottlenecks in backward modeling. Bayesian inference has typically been positioned to form the core of a DT instantiation’s decision engine; however, forming a posterior with high-fidelity hydrodynamic forward models is computationally taxing. Agentic AI and ML technologies can help, effectively shifting the burden of contextualization from Bayesian inference to more data-driven technologies. Herein, we demonstrate one such alternative for the purposes of correcting forward models on-the-fly, the transformer. The transformer, which underlies the popular large language models (LLMs) of today, is shown to attend to myriad exogenous data and generalize across a range of hydrodynamic conditions, making it an ideal candidate for rapid forward model improvement. We demonstrate its use on NOAA’s Global Storm Surge and Tide Forecast System (STOFS-2D-Global). Additionally, for applications where we cannot bypass Bayesian inference or otherwise need to propagate uncertainty quickly, we demonstrate a low-fidelity predictor that, when coupled to performant forward models like STOFS, could be used to efficiently estimate the posterior. This predictor is shown to generalize across a range of complex coastal geometries for multi-day forecasts. These two data-driven examples serve to motivate agentic AI (e.g. model context protocol-capable LLMs) for DT instantiations for coastal flooding.