Multidecadal Oceanic Modulation of Summer Precipitation in North China in 1200-Year Global Climate Simulations

• Published in: Journal of climate Vol. 36; no. 17; pp. 6125 - 6138
• Main Authors: Tong, Xuan, Yan, Zhongwei, Zhou, Wen, Xia, Jiangjiang, Quan, Xiaowei
• Format: Journal Article
• Language: English
• Published: 01.09.2023
• ISSN: 0894-8755; 1520-0442
• DOI: 10.1175/JCLI-D-22-0693.1

Abstract The multidecadal decrease of precipitation in North China since the 1960s has caused a significant impact on the ecological environment and social development. It has become a key concern for meteorologists and national decision-makers to understand what is driving this. To better understand the oceanic impact on the interdecadal variation of summer precipitation in North China, we analyze two 1200-yr global climate simulations from state-of-the-art coupled models with an explainable convolutional neural network (CNN). The accuracy of the CNN model in correctly diagnosing the increased and decreased summer precipitation anomaly in North China reaches 98% and 96.6% in two CMIP6 models, respectively. We find that a sea surface temperature anomaly (SSTA) in the northern Pacific, associated with the Pacific decadal oscillation (PDO), and an SSTA in the northern Atlantic, associated with the Atlantic multidecadal oscillation (AMO), play essential roles in the interdecadal variation of summer precipitation in North China. The modulations of the PDO and AMO on summer precipitation in North China are consistent with the results derived from limited observations and simulations, which means these mechanisms are robust. In particular, when the PDO is in a negative phase and the AMO is in a positive phase, the positive anomaly of precipitation and the associated circulation anomalies are intensified. This study also confirms that with the explainable deep learning method, a convolutional neural network can be used to identify multidecadal variability in North China summer precipitation, although limited datasets such as observational records make it more challenging to apply deep learning methods.