相对于单一极端天气气候事件，由极端强风和极端降水造成的复合极端风雨事件（Compoundwind and precipitation extremes，CWPE）对沿海地区的经济和人民生活造成巨大的影响。本文利用第六次国际耦合模式比较计划（Coupled Model Intercomparison Project Phase 6，CMIP6）中 17 个模式模拟1961-2000 年中国东部沿海地区 CWPE 作为训练集，采用多层感知机神经网络建立深度学习（DeepLearning，DL）模型。通过构建适用于 CWPE损失函数的方式对模型进行优化，发展适用于降低 CMIP6模式对CWPE模拟偏差和不确定性的DL模型。研究结果表明，大多数CMIP6模式对中国东部沿海地区CWPE有较好的模拟能力，其中多模式集合平均值（Multi-Model Ensemble Mean，MME-Mean）和多模式集合中位数（Multi-Model Ensemble Median，MME-Median）的结果相对于单一模式的评估表现更好。以均方误差（Mean Squared Error，MSE）函数作为损失函数构建的 DL 模型，在泰勒技巧评分（Taylor SkillScore，TS）和均方根误差（Root Mean Squared Error，RMSE）评分上的表现不如多模式集合统计结果。在MSE损失函数基础上加入气候评估指标中的标准差之比（Ratio of the Standard Deviation，RSD）和构建的低估值约束函数，可以有效提高DL模型在TS和RMSE评分上的表现。因此，将由MSE、RSD以及低估值约束函数构建的加权损失函数训练的DL模型定义为DL-MRM，仅以MSE作为损失函数训练的DL模型定义为 DL-MSE。对比分析 2001-2014 年间两个 DL 模型对中国东部沿海 CWPE 模拟表现以及 DLMRM相对于多模式集合方法的表现得出：（1）两个DL模型模拟结果均表现低估，但DL-MRM的偏差相比于 DL-MSE 更低且更接近观测，其中在研究区域内 DL-MRM 的相对偏差低于 DL-MSE 的面积约为63%，且相对偏差平均降低了约20%；（2）DL-MRM相较于MME-Mean和MME-Median，整体偏差较低，其模拟结果更接近观测，在研究区域内 DL-MRM 的相对偏差较低的面积占比分别为 67%和 62%，且相对偏差分别平均降低约 10%和20%。总体而言，通过融合 RSD和低估值约束函数构建加权损失函数的方式对模型进行优化，建立了适用于提高CMIP6模式模拟CWPE能力的DL模型，并表明结合深度学习方法相对于传统多模式集合方法能更有效地降低CMIP6模式模拟CWPE的偏差。

Relative to individual extreme weather and climate events，Compound Wind and Precipitation Extremes（CWPE），which result from extreme winds and extreme precipitation，have a profound impact on the economy and daily life in coastal areas. In this study，we utilized the simulations from 17 models within the Coupled Model Intercomparison Project Phase 6（CMIP6）for the period 1961-2000 of CWPE in the eastern coastal region of China as a training set，and established a Deep Learning（DL）model employing a multi-layer perceptron neural network. By constructing a loss function suitable for CWPE and optimizing the model accordingly， we have developed a DL model aimed at reducing the simulation bias and uncertainty of CMIP6 models for CWPE. The research results indicate that the majority of CMIP6 models possess a relatively good simulation capability for CWPE in the eastern coastal region of China，with the Multi-Model Ensemble Mean（MME-Mean） and the Multi-Model Ensemble Median（MME-Median）demonstrating better performance in assessments com‐ pared to individual models. The DL model constructed with the Mean Squared Error（MSE）function as the loss function performs worse in terms of Taylor Skill Score（TS）and Root Mean Squared Error（RMSE）compared to the statistical results of the Multi-Model Ensemble. Incorporating the Ratio of the Standard Deviation（RSD） from climate evaluation metrics and an underestimation constraint function into the MSE loss function can significantly enhance the performance of the DL model in terms of TS and RMSE. Therefore，the DL model trained with a weighted loss function constructed from MSE，RSD，and underestimation constraint function is defined as DL-MRM，while the DL model trained solely with MSE as the loss function is defined as DL-MSE. By com‐ paring and analyzing the performance of the two DL models in simulating CWPE over the eastern coastal region of China from 2001 to 2014，as well as the performance of DL-MRM relative to multi-model ensemble methods， we conclude：（1）Both DL models exhibit underestimation in their simulation results，but the bias of the DLMRM is lower than that of the DL-MSE，being closer to the observations. Specifically，in the study area，the relative bias of the DL-MRM is lower than that of the DL-MSE by about 63%，and the average relative bias is reduced by approximately 20%.（2）The DL-MRM has a lower overall bias compared to the MME-Mean and MME-Median，with simulation results that are closer to the observations. In the study area，the DL-MRM has a lower relative bias in 67% and 62% of the area compared to the MME-Mean and MME-Median，respectively， and the average relative bias is reduced by approximately 10% and 20%，respectively. Overall，by integrating the RSD and underestimation constraint functions to construct a weighted loss function for model optimization，a DL model suitable for improving the simulation of CWPE by CMIP6 models was established. This indicates that the combination of deep learning methods can more effectively reduce the biases in CMIP6 model simulations of CWPE compared to traditional multi-model ensemble methods.