A hybrid physics-informed graph neural network for tornado wind-field modeling

The accurate prediction of tornado wind fields is paramount for the wind-resistant design of critical infrastructure. Analytical models offer a computationally efficient and physically grounded foundation for representing the primary structure of tornadic vortices. While these models provide valuable approximations, they inherently simplify the highly complex turbulent characteristics observed in tornado phenomena. Conversely, high-fidelity methods like Large-Eddy Simulations (LES) capture these intricate details with high accuracy but are computationally prohibitive for many direct engineering applications. This study introduces a novel hybrid framework that integrates a physics-based analytical model with a physics-informed deep learning model to bridge this gap. The proposed approach incorporates physics in two distinct steps. First, an established analytical model (the Baker model) is utilized to generate a baseline velocity field, providing a robust physical approximation of the vortex. Second, a Graph Neural Network (GNN) is trained to learn the complex, non-linear residuals between this analytical baseline and ground-truth LES data sourced from foundational numerical studies. The GNN architecture is designed to capture spatial interdependencies within the vortex field, with training guided by physics-based penalty functions to ensure physical consistency. The results demonstrate that the hybrid GNN model achieves a global test-set root-mean-square error (RMSE) of 0.0039, a substantial improvement compared to the analytical baseline’s test-set RMSE of 0.8384. This framework presents a computationally efficient alternative that preserves the high fidelity of LES mean flow structures, offering a promising tool for advanced engineering analysis and design.