Improving CMIP6 Atmospheric River Precipitation Estimation by Cycle-Consistent Generative Adversarial Networks

Given the important role of Atmospheric River precipitation (ARP) in the global hydrological cycle, accurate representation of ARP is significant. However, general circulation models (GCMs) demonstrate bias in simulating ARP. The target of this study is to quantify the performance of ARP intensity/frequency for CMIP6 simulations, and further to improve ARP estimation of CMIP6 using Cycle-Consistent Generative Adversarial Networks (CycleGAN) with highlighting the more accurate ARP features under the global warming background. The findings of this study are as follows: (a) although ARP intensity/frequency in reserved-optimal CMIP6 overall reproduces the observation, it is still underestimated at the stronger Atmospheric river (AR) scales, particularly for the AR highly active mid-latitude regions. (b) The CycleGAN-based bias correction approach markedly diminishes the bias of the CMIP6 simulations within most of the AR scales among both global and the four AR highly active regions. Moreover, the performance of the ARP in AR highly active regions is significant improvement, which is mainly due to the reduction of the bias at the strongest scale. (c) Relative to reference period (1986-2005), ARP intensity/frequency at the strongest scale increase notably under 3 degrees C warming level, with an average value of 373.3% in intensity and 415.9% in frequency for global and the four key regions before correction, and the value is 451.9% and 492.5% after bias correction. The results illustrate that CycleGAN can effectively improve the ARP simulations of GCMs, and an early warning implies that future strong extreme ARP should potentially surpass the current expected. Despite the important role of Atmospheric River precipitation (ARP) in global hydrological cycle, accurately simulating global ARP remains a great challenge. Here, the content of our study consists of three main parts. Firstly, we quantify the performance of CMIP6 general circulation models (GCMs) on ARP intensity/frequency simulations and find that CMIP6 GCMs underestimates ARP at the Atmospheric river (AR) highly active regions, especially in the strongest AR intensity scale. Secondly, we utilize a deep learning method named CycleGAN to improve global ARP intensity/frequency simulations of CMIP6 GCMs. The results indicate that CycleGAN effectively corrects the ARP intensity/frequency bias within most of the AR intensity scales among both global and the AR highly active regions. Moreover, the performance of ARP in the AR highly active regions is significant improvement, which is mainly due to the reduction of the bias in the strongest scale. Finally, the trained CycleGAN model is used to evaluate ARP under global warming of 1.5, 2, and 3 degrees. The bias correction results show that ARP in the strongest AR intensity scale surpasses the current expected, especially under 3 degrees warming level. CMIP6 models underestimates Atmospheric River precipitation (ARP) at the Atmospheric river (AR) highly active regions, especially in the strongest AR scale The CycleGAN-based deep learning method reduces model bias notably for ARP intensity and frequency, respectively ARP in the strongest AR scale potentially surpasses the current GCM-based projections expected in different degrees of warming after bias correction