This study develops a machine learning (ML)-based physics parameterization suite trained on 80 d global storm-resolving model (GSRM) simulation data (5 km), with the aim of replacing all conventional physics tendencies in a general circulation model (GCM, 120 km) for real-world simulations with realistic surface topography. The GSRM data are generated using the Global–Regional Integrated Forecast System (GRIST) and subsequently coarse-grained, after which the residual method is applied to derive the corresponding GCM physics tendencies. The resulting workflow relies on standardized pressure-level variables as input features, enabling the GCM – through physics–dynamics coupling – to effectively emulate the multiscale flow interactions captured by the GSRM. This ML-enhanced GCM sustains stable 6-year Atmospheric Model Intercomparison Project (AMIP) type simulations and produces a realistic climatology comparable to that of a skillful GCM. It effectively mitigates the biases of excessively strong rainbands and an overly wide ITCZ in the conventional configuration, when compared with the Global Precipitation Measurement (GPM) data. Moreover, the hybrid ML-GCM better captures precipitation frequency, notably mitigating the overproduction of light tropical rainfall. Sensitivity experiments using different neural network architectures (ResNet, CNN, MLP) demonstrate that all configurations can maintain long-term simulation stability, with ResNet showing superior simulation accuracy. This work presents a transferable framework that leverages km-scale GSRM data to enhance GCM performance via ML integration, offering a potential route to reduce the gaps between two modeling paradigms.

多尺度模拟，全球风暴解析模拟，全球气候模拟，,人工智能物理参数化研究