Reconstruction of tropical cyclone boundary layer wind field using physics-informed machine learning

Abstract

A physics-informed machine learning model is proposed in this paper to reconstruct the high-fidelity three-dimensional boundary layer wind field of tropical cyclones. The governing equations of the wind field, which incorporate a spatially varying eddy diffusivity coefficient, are derived and embedded within the model's loss function. This integration allows the model to learn the underlying physics of the boundary layer wind field. The model is applied to reconstruct two tropical cyclone events in different oceanic basins. A wide range of observational data from satellite, dropsonde, and Doppler radar records are assimilated into the model. The model's performance is evaluated by comparing its results with observations and a classic linear model. The findings demonstrate that the model's accuracy improves with an increased amount of real data and the introduction of spatially varying eddy diffusivity. Furthermore, the proposed model does not require strict boundary conditions to reconstruct the wind field, offering greater flexibility compared to traditional numerical models. With the assimilation of observational data, the proposed model accurately reconstructs the horizontal, radial, and vertical distributions of the wind field. Compared with the linear model, the proposed model more effectively captures the nonlinearities and asymmetries of the wind field, thus presents more realistic outcomes.