Leveraging Scientific AI to Accelerate Simulation-Based Natural Hazard Risk Assessment and Forecasting

Abstract

Accurate modeling of coastal hazards, including tropical cyclones (TCs) and storm surges, is critical for risk assessment and enhancing community resilience. This research introduces a suite of advanced deep learning frameworks designed to address key limitations in current hazard modeling.

For expanding the historical TC database, we leverage two deep generative models. A diffusion-based approach, TC-Diffusion, generates realistic full TC tracks by inherently processing spatial heterogeneity, avoiding the need for segmentation. Complementing this, a flow-based model simulates environment-dependent TC intensity evolution, outperforming traditional stochastic models. Both methods demonstrate excellent agreement with historical data, proving effective for wind hazard assessment.

For storm surge modeling, two research questions are tackled. To improve real-time fore-casting, we propose a spatio-temporal framework that integrates Graph Neural Networks (GNN) and Gated Recurrent Units (GRU). This model captures complex, causality-informed dependencies between observation stations, significantly outperforming baseline models in short-term predictions. For efficient long-term risk assessment, we developed Surge-NF, a novel surrogate model inspired by Neural Fields. By using positional encoding and a multi-task learning framework to predict both peak surge and dry-wet status, Surge-NF overcomes the over-smoothing and data-inefficiency of existing surrogates, drastically reducing computational cost and error.

Collectively, these studies demonstrate the power of specialized deep learning architectures to create more accurate, efficient, and physically informed hazard models. By overcoming challenges such as spatial heterogeneity, inter-site dependencies, and computational expense, these frameworks significantly advance our capabilities in TC simulation and storm surge forecasting, enabling better risk management and safeguarding coastal communities.

Date of Award	2025
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology
Supervisor	Jize ZHANG (Supervisor) & Tim K.t. Tse (Supervisor)