大气环流模式（GCM）自身分辨率难以达到云分辨尺度，采用次网格参数化方案描述云和对流等湿物理过程。然而这些参数化方案对于云和对流描述不精确，造成模式在气候态和内部变率模拟上存在偏差，对未来气候的预估也存在不确定性。数据科学的快速发展，构建机器学习的物理参数化方案成为可能。本研究首次采用深度卷积残差神经网络的先进结构，以超级参数化方案（SPCAM）的模拟结果为训练数据，在国际上率先成功实现在具有真实海陆分布和地形的气候模式中用神经网络模拟对流及云的湿物理过程，并具有高精度的离线模拟和多气候态预测能力。该方案可单柱模式中完成海洋对流过程和陆地强降水预报，并且在三维大气环流模式耦合模拟中有5天精确的表现。本研究创新性的考虑了历史对流记忆对当前对流预报的影响；在网络预报变量中加入了质量守恒，在训练的损失函数中加入了湿静力能守恒。结合深度学习可视化的方法，本研究首次对深度学习方案进行系统的物理解释，排序了预报因子的重要性，探究了内部物理逻辑，探索了深度学习方案的可解释性。本研究确认了对流记忆变量为最重要的因子，并且总结了在深对流背景条件下的线性响应函数。此外，本研究探索了网络结构与线性响应函数之间的关系，包括线性函数和简化二维重力波方程与深度学习参数化方案稳定性的相关性。在与计算机系合作开发长期稳定模拟深度学习参数化方案的过程中，本研究将替代的范围由湿物理扩大到湿物理和辐射过程，由此确定了预报因子和预报变量。同时，本研究还评估了该参数化方案在线5年的模拟结果：其多年气候平均的温度和水汽存在一定误差，但是可以再现SPCAM的极端降水发生频率和季节内振荡。
General circulation models （GCMs） are important means to simulate and study climate changes. However, due to the inaccurate description of cloud and convection in GCMs, there are not only deviations in the simulation of climatology means and internal variability, but also a high degree of uncertainty in future climate projections. Because the resolution of GCMs cannot reach the cloud scale, using the traditional cumulus convection parameterizations to describe the cloud and convection process is a large source of simulation error.The rapid development of data science makes it possible to build machine learning based parameterizations. This study will explore deep learning algorithms on moist physics parameterizations under realistic configurations. The simulations of a Superparameterized GCM （SPCAM） in realistic configurations are used as training data. This study first applied the deep residual convolutional neural network （ResNet）, which has higher parameter utilization efficiency and stronger nonlinear fitting capability. Its inputs are the GCM grid-scale environmental states, the large-scale forcings, and the convective memory variables. And it predicts heating rate and moistening rate and cloud properties in moist physics. Furthermore, we innovatively consider the effects of convective history on new convections by taking the SPCAM’s convection and cloud properties at previous timesteps as input variables. We also add a penalty term to the loss function to balance the prediction accuracy and conservation of moist static energy. This ResNet can reduplicate the SPCAM target results in the independent testing dataset, draw more similarities to the observations in a single column model （SCM）, and achieve a short period of high prognostic performance in a 3D GCM. In addition, it carries out a good prediction in a warm climate.This study applies machine learning visualization methods to explore the main factors and potential physical laws behind the deep learning parametrizations. The importance of convective memory is confirmed by ranking all considered input variables. This study also summarizes the linear response function （LRF） of deep learning parameterizations and tests their stability by combining the LRF with a simplified 2D gravity wave system.This study collaborated with the computer science department to develop a stable machine learning parameterization. During the process, this study expands the moist-physics deep learning parameterization to include the radiation process, thereby redesigning the predictors and predictors. The new parameterization can achieve stable simulation for at least 5 years. In the 5-year prognostic results, even with some deviations in the climate mean temperature and humidity, the deep learning enabled GCM can reproduce the precipitation probability distribution and MJO propagation speed in SPCAM.
