A physics-informed neural network-based approach to reconstruct the tornado vortices from limited observed data

Abstract

The measured wind data of tornado vortices is often insufficient and low-resolution for analysis of tornado-induced loads and effects on civil structures. This study proposes a physics-informed neural network-based (PINN-based) approach to reconstruct the tornado vortices from limited observed data. First, the PINN model is constructed by embedding the simplified tornado governing equations in a deep neural network as the prior physical information. The PINN model is then validated by the reconstruction of different numerical tornado vortices generated by large eddy simulations. Parametric analysis is also conducted to investigate the influence of the structure of the PINN and the number of observation points. Finally, the proposed approach is applied to reconstruct a real tornado vortex. Results show that the PINN-based method can accurately reconstruct the numerical tornado vortices in different stages with a relative error below 4% for tangential velocity and vertical velocity, and 11% for radial velocity. Increasing the number of observation points and the number of points for training physical equations will improve the reconstruction accuracy. The proposed approach can correctly reconstruct the real tornado vortex with a relative error of less than 10%.