Search

In climate models, subgrid parameterizations of convection and clouds are one of the main causes of the biases in precipitation and atmospheric circulation simulations. In recent years, due to the rapid development of data science, machine learning (ML) parameterizations for convection and clouds have been demonstrated to have the potential to perform better than conventional parameterizations. Most previous studies were conducted on aqua-planet and idealized models, and the problems of simulation instability and climate drift still exist. Developing an ML parameterization scheme remains a challenging task in realistically configured models. In this paper, a set of residual deep neural networks (ResDNNs) with a strong nonlinear fitting ability is designed to emulate a super-parameterization (SP) with different outputs in a hybrid ML–physical general circulation model (GCM). It can sustain stable simulations for over 10 years under real-world geographical boundary conditions. We explore the relationship between the accuracy and stability by validating multiple deep neural network (DNN) and ResDNN sets in prognostic runs. In addition, there are significant differences in the prognostic results of the stable ResDNN sets. Therefore, trial and error is used to acquire the optimal ResDNN set for both high skill and long-term stability, which we name the neural network (NN) parameterization. In offline validation, the neural network parameterization can emulate the SP in mid- to high-latitude regions with a high accuracy. However, its prediction skill over tropical ocean areas still needs improvement. In the multi-year prognostic test, the hybrid ML–physical GCM simulates the tropical precipitation well over land and significantly improves the frequency of the precipitation extremes, which are vastly underestimated in the Community Atmospheric Model version 5 (CAM5), with a horizontal resolution of 1.9∘ × 2.5∘. Furthermore, the hybrid ML–physical GCM simulates the robust signal of the Madden–Julian oscillation with a more reasonable propagation speed than CAM5. However, there are still substantial biases with the hybrid ML–physical GCM in the mean states, including the temperature field in the tropopause and at high latitudes and the precipitation over tropical oceanic regions, which are larger than those in CAM5. This study is a pioneer in achieving multi-year stable climate simulations using a hybrid ML–physical GCM under actual land–ocean boundary conditions that become sustained over 30 times faster than the target SP. It demonstrates the emerging potential of using ML parameterizations in climate simulations.

A stochastic deep convection parameterization is implemented into the US Department of Energy (DOE) Energy Exascale Earth System Model (E3SM) Atmosphere Model version 1.0 (EAMv1). This study evaluates its performance in simulating precipitation. Compared to the default model, the probability distribution function (PDF) of rainfall intensity in the new simulation is greatly improved. The well-known problem of “too much light rain and too little heavy rain” is alleviated, especially over the tropics. As a result, the contribution from different rain rates to the total precipitation amount is shifted toward heavier rain. The less frequent occurrence of convection contributes to suppressed light rain, while more intense large-scale and convective precipitation contributes to enhanced heavy total rain. The synoptic and intraseasonal variabilities of precipitation are enhanced as well to be closer to observations. The sensitivity of the rainfall intensity PDF to the model vertical resolution is examined. The relationship between precipitation and dilute convective available potential energy in the stochastic simulation agrees better with that in the Atmospheric Radiation Measurement (ARM) observations compared with the standard model simulation. The annual mean precipitation is largely unchanged with the use of the stochastic scheme except over the tropical western Pacific, where a moderate increase in precipitation represents a slight improvement. The responses of precipitation and its extremes to climate warming are similar with or without the stochastic deep convection scheme.

The Energy Exascale Earth System Model (E3SM) developed by the Department of Energy has a goal of addressing challenges in understanding the global water cycle. Success depends on correct simulation of cloud and precipitation elements. However, lack of appropriate evaluation metrics has hindered the accurate representation of these elements in general circulation models. We derive metrics from the three-dimensional data of the ground-based Next-Generation Radar (NEXRAD) network over the US to evaluate both horizontal and vertical structures of precipitation elements. We coarsened the resolution of the radar observations to be consistent with the model resolution and improved the coupling of the Cloud Feedback Model Intercomparison Project Observation Simulator Package (COSP) and E3SM Atmospheric Model Version 1 (EAMv1) to obtain the best possible model output for comparison with the observations. Three warm seasons (2014–2016) of EAMv1 simulations of 3-D radar reflectivity features at an hourly scale are evaluated. A general agreement in domain-mean radar reflectivity intensity is found between EAMv1 and NEXRAD below 4 km altitude; however, the model underestimates reflectivity over the central US, which suggests that the model does not capture the mesoscale convective systems that produce much of the precipitation in that region. The shape of the model-estimated histogram of subgrid-scale reflectivity is improved by correcting the microphysical assumptions in COSP. Different from previous studies that evaluated modeled cloud top height, we find the model severely underestimates radar reflectivity at upper levels – the simulated echo top height is about 5 km lower than in observations – and this result is not changed by tuning any single physics parameter. For more accurate model evaluation, a higher-order consistency between the COSP and the host model is warranted in future studies.

Changes in the large-scale environment during convective precipitation events in the tropical western Pacific simulated by version 4.3 of the Canadian Atmospheric Model (CanAM4.3) are compared against those simulated by version 5.0 of the super-parameterized Community Atmosphere Model (spCAM5). This is done by compositing sub-hourly output of convective rainfall, convective available potential energy (CAPE), CAPE generation due to large-scale forcing in the free troposphere (dCAPELSFT) and near-surface vertical velocity (ω) over the time period May–July 1997. Compared to spCAM5, CanAM4.3 tends to produce more frequent light convective precipitation ( mm h−1) and underestimates the frequency of extreme convective precipitation (>2 mm h−1). In spCAM5, 5 % of convective precipitation events lasted less than 1.5 h and 75 % lasted between 1.5 and 3.0 h, while in CanAM4.3 80 % of the events lasted less than 1.5 h. Convective precipitation in spCAM5 is found to be a function of dCAPELSFT and the large-scale near-surface ω with variations in ω slightly leading variations in convective precipitation. Convective precipitation in CanAM4.3 does not have the same dependency and instead is found to be a function of CAPE.