Multi-fidelity graph neural networks for efficient and accurate flood hazard mapping

Abstract

Generating high-resolution flood hazard maps with traditional hydrodynamic models is computationally prohibitive. While surrogate models like graph neural networks (GNNs) offer a faster alternative, they require large, expensive-to-generate, high-fidelity training datasets. This study addresses the critical challenge of creating accurate surrogate models with limited high-fidelity data. We propose a novel multi-fidelity graph neural network (MFGNN) framework that integrates numerous inexpensive, coarse-resolution simulations with a few high-fidelity runs. The method uses a hierarchical pipeline where one GNN learns broad flood patterns from low-fidelity data, and a second GNN learns to predict and apply a high-resolution correction based on the residual error. Comprehensive validation across diverse fluvial and pluvial flood scenarios demonstrates that the MFGNN framework significantly reduces prediction error compared to a standard GNN trained with an equivalent computational budget. This computationally efficient, novel framework makes development of accurate, high-resolution surrogate models for large floodplains feasible by lowering the data-generation barrier.