Stable climate simulations using a realistic general circulation model with neural network parameterizations for atmospheric moist physics and radiation processes

Abstract. 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.