Physics Informed Machine Learning for Rapid and Enhanced Tropical Cyclone Forecasts

Tropical cyclones (TC) have historically been the most damaging of weather disasters—a trend that is expected to worsen because of sea level rise, land subsidence, and the intensification of TCs driven by climate change. Traditionally, storm surge forecasting relies on numerical models, such as the National Oceanic and Atmospheric Administration's (NOAA) Sea, Lake, and Overland Surges from Hurricanes (SLOSH). As the SLOSH model runs in minutes, it is utilized by the Probabilistic Tropical Cyclone Storm Surge (P-Surge) model, to create probabilistic ensembles of TC-induced water depths based on different storm paths and intensities. However, while SLOSH is efficient in simulating 1,000s of ensemble members, it does not account for wave effects and lacks the high spatial resolution necessary for accurate overland flooding assessment. In contrast, high-resolution storm surge models, such as ADCIRC, offer greater accuracy but are computationally expensive, making them unable to simulate the 1,000s of ensemble members quickly enough for forecasting. This presents a challenge: how to leverage the accuracy of high-fidelity models while achieving the efficiency of low-fidelity models. To address this challenge, we demonstrate a multi-fidelity modeling approach where machine learning (ML) corrects low-fidelity ADCIRC model outputs (simulated in minutes), aligning them with high-fidelity ADCIRC results. The tree-based ML model learns these corrections by considering joint probability method (JPM) parameters (e.g., forward velocity, central pressure deficit), along with local (point-specific) and nonlocal (regional) characteristics. This approach allows for the rapid generation of high-fidelity tropical cyclone water levels that closely match those produced by high-resolution ADCIRC modeling, thereby enabling the efficient simulation of probabilistic ensembles based on the uncertainty in the forecasted TC tracks and intensities. Accordingly, we anticipate that this approach may enhance existing tropical cyclone forecasting methods (e.g., P-Surge) through the creation of more accurate ensembles of probabilistic TC water depths over inland areas.