Training data from SPCAM for machine learning in moist physics

Current moist physics parameterization schemes in general circulation models (GCMs) are the main source of biases in simulated precipitation and atmospheric circulation. Recent advances in machine learning make it possible to explore data-driven approaches to developing parameterization for moist physics processes such as convection and clouds. This study aims to develop a new moist physics parameterization scheme based on deep learning. We use a residual convolutional neural network (ResNet) for this purpose. It is trained with one-year simulation from a superparameterized GCM, SPCAM. An independent year of SPCAM simulation is used for evaluation. In the design of the neural network, referred to as ResCu, the moist static energy conservation during moist processes is considered. In addition, the past history of the atmospheric states, convection and clouds are also considered. The predicted variables from the neural network are GCM grid-scale heating and drying rates by convection and clouds, and cloud liquid and ice water contents. Precipitation is derived from predicted moisture tendency. In the independent-data test, ResCu can accurately reproduce the SPCAM simulation in both time-mean and temporal variance. Comparison with other neural networks demonstrates the superior performance of ResNet architecture. ResCu is further tested in a single column model for both continental midlatitude warm season convection and tropical monsoonal convection. In both cases, it simulates the timing and intensity of convective events well. In the prognostic test of tropical convection case, the simulated temperature and moisture biases with ResCu are smaller than those using conventional convection and cloud parameterizations.

Methods
This dataset is extracted from a simulation using a Superparameterized GCM, SPCAM (https://wiki.ucar.edu/display/ccsm/Superparameterized+CAM+(SPCAM)). The SPCAM implements a 2-D CRM in CAM5.2 to replace its conventional parameterization for moist convection and large-scale condensation. The dynamic framework of CAM5 has a horizontal resolution of 1.9x2.5 degrees and 30 vertical levels that are shared with the embedded CRM. The SPCAM used in this study has a coupled land surface model Community Land Model 4.0 (CLM4.0). It uses a prescribed climatological sea surface temperature field that comes with the CAM5 model. It is run for three years and 4 months from Jan. 1st in 1998 to March 31st in 2001 with a time step of 20 minutes. The first year and 4 months are for spin up, the second year is used for training and the third year is used for testing and evaluation. The training data from SPCAM is output every timestep. This dataset contains one year training data and one year evaluation data.

通用环流模式（General Circulation Models，GCM）中现有的湿物理参数化方案，是模拟降水与大气环流结果产生偏差的主要来源。近年来机器学习领域的进展，为探索基于数据驱动的湿物理过程参数化方案开发路径提供了可能，这类湿物理过程涵盖对流与云系过程。本研究旨在开发一种基于深度学习的新型湿物理参数化方案，为此采用了残差卷积神经网络（Residual Convolutional Neural Network，ResNet）。该模型以超参数化通用环流模式（Superparameterized GCM, SPCAM）产出的一年模拟数据作为训练集，并采用另一独立年份的SPCAM模拟结果开展评估。在这款被命名为ResCu的神经网络设计中，研究团队考虑了湿过程中的湿静能守恒特性。此外，该网络还纳入了大气状态、对流活动与云系的历史演变信息。该神经网络的预测变量包括：对流与云系引发的GCM格点尺度加热、干燥速率，以及云液态水与冰态水含量。降水结果则由预测得到的水汽倾向推导得出。在独立数据集测试中，ResCu能够准确复现SPCAM的模拟结果，无论是时间平均特征还是时间变率特征。与其他神经网络的对比实验表明，ResNet架构具备更优异的性能表现。研究团队进一步将ResCu应用于单柱模式中，分别针对大陆中纬度暖季对流与热带季风对流开展测试。在两类测试场景中，该模型均能准确模拟对流事件的发生时机与强度。在热带对流案例的预报试验中，ResCu模拟得到的温度与水汽偏差，小于采用传统对流与云参数化方案的模拟结果。

方法
本数据集提取自超参数化通用环流模式SPCAM（https://wiki.ucar.edu/display/ccsm/Superparameterized+CAM+(SPCAM)）的模拟结果。SPCAM在CAM5.2模式中嵌入了二维云解析模型（Cloud Resolving Model, CRM），以替代原有的湿对流与大尺度凝结传统参数化方案。CAM5的动力框架水平分辨率为1.9°×2.5°，共设置30个垂直层，该设置与内嵌的CRM保持一致。本研究使用的SPCAM耦合了陆面模式通用陆面模式4.0（Community Land Model 4.0, CLM4.0）。该模式采用CAM5模式自带的指定气候学海表温度场。模式模拟时长为3年4个月，自1998年1月1日起至2001年3月31日止，时间步长设为20分钟。前1年4个月用于模式自旋调整阶段，第二年作为训练数据集，第三年用于测试与评估。SPCAM的训练数据按每个时间步输出。本数据集包含1年的训练数据与1年的评估数据。