Deep Learning Parameterization of the Tropical Cyclone Boundary Layer.

The research on tropical cyclone (TC) relieson numerical models and simulations, with the currently widely used boundary layer parameterization posing a significant challenge on accurately predicting turbulent mixing. However, machine learning has opened up new possibilities for boundary layer sub‐grid process parameterization. In this study, a deep‐learning parameterization scheme for the TC boundary layer (DeepBL) on predicting turbulent flux is proposed. DeepBL comprises a one‐dimensional convolutional structure that relies on a small number of learnable parameters that accomplishes an error reduction compared to the standard fully connected neural networks. Furthermore, a nonlinear transformation scheme is introduced to alleviate the training data's skewness and improve the DeepBL performance by affording a smaller prediction error. Specifically, the output data of a large‐eddy simulation of an idealized TC are used to train, validate, and test DeepBL, affording significantly better performance than the YSU scheme in the Weather Research and Forecast model. Interpretability analysis on DeepBL demonstrates that the deep‐learning parameterization scheme is physically reasonable. Plain Language Summary: Traditional boundary layer parameterizations explicitly program algorithms to reproduce the effects of turbulent processes in numerical models. They suffer from evident biases when simulating the tropical cyclone boundary layer. Therefore, a deep learning model with a concise structure is introduced to improve the boundary‐layer representation. The developed model is trained on the output of a large‐eddy simulation of a tropical cyclone, where the turbulent processes are explicitly simulated. Although the unbalanced training data distribution will hinder the deep learning model's performance, the nonlinear transformation proposed in this study assists the training data distribution to be more compact and reduce the prediction errors of the trained model. The results demonstrate that the proposed deep‐learning model is significantly improved compared with traditional methods in reproducing the spatial features and data distribution of turbulent fluxes. Key Points: A neural network (NN) with a one‐dimensional convolutional structure is designed to parameterize the tropical cyclone boundary layerThe proposed NN can reproduce the spatial features and data distribution of turbulent fluxes better than the traditional schemeA nonlinear transformation applied to the training data of the NN can markedly reduce errors for turbulent flux prediction [ABSTRACT FROM AUTHOR]