Journal Article

Stable Machine‐Learning Parameterization of Subgrid Processes in a Comprehensive Atmospheric Model Learned From Embedded Convection‐Permitting Simulations

Hu, Zeyuan, Subramaniam, Akshay, Brenowitz, Noah D., Kuang, Zhiming, Pritchard, Michael S., Lin, Jerry, Yu, Sungduk, Hannah, Walter M., Romero, Josh

2025

Overview

Modern climate projections often suffer from inadequate spatial and temporal resolution due to computational limitations, resulting in inaccurate representations of sub‐grid processes. A promising technique to address this is the multiscale modeling framework (MMF), which embeds a kilometer‐resolution cloud‐resolving model (CRM) within each atmospheric column of a host climate model to replace traditional convection and cloud parameterizations. Machine learning offers a unique opportunity to make MMF more accessible by emulating the embedded CRM and reducing its substantial computational cost. Although many studies have demonstrated proof‐of‐concept success of achieving stable hybrid simulations, it remains a challenge to achieve near operational‐level success with real geography and comprehensive variable emulation that includes, for example, explicit cloud condensate coupling. In this study, we present a stable hybrid model capable of integrating for at least 5 years with near operational‐level complexity, including coarse‐grid geography, seasonality, explicit cloud condensate and wind predictions, and land coupling. Our model demonstrates skillful online performance, achieving a 5‐year zonal mean tropospheric temperature bias within 2 K, water vapor bias within 1 g/kg, and a precipitation root mean square error of 0.96 mm/day. Key factors contributing to our online performance include an expressive U‐Net architecture and physical thermodynamic constraints for microphysics. With microphysical constraints mitigating unrealistic cloud formation, our work is the first to demonstrate realistic multi‐year cloud condensate climatology under the MMF framework. Despite these advances, online diagnostics reveal persistent biases in certain regions, highlighting the need for innovative strategies to further optimize online performance. Plain Language Summary Traditional climate models often struggle to accurately simulate small‐scale processes like thunderstorms due to compute limitations, leading to less reliable climate predictions. Machine learning (ML) offers a promising solution by efficiently modeling these processes and integrating them into hybrid ML‐physics simulations within a host climate model. While previous studies have shown success in simplified setups, such as all‐ocean planets, achieving accurate and stable hybrid simulations in complex, real‐world settings remains challenging. In this study, we developed a stable hybrid model capable of simulating the climate for 5 years using real geographic features and explicitly predicting the time evolution of temperature, moisture, cloud, and wind. Our model achieves skillful accuracy in long‐term mean atmospheric states. This success is due to several key improvements: an advanced architecture and the incorporation of cloud physics constraints. Key Points Stable hybrid climate simulations are achieved with a data‐driven emulator of subgrid physics coupled with a comprehensive atmosphere model Online performance benefits from a U‐Net architecture and microphysical constraints A realistic cloud climatology with explicit cloud condensate coupling is achieved in a hybrid multi‐scale modeling framework

• Publisher: John Wiley & Sons, Inc,American Geophysical Union (AGU)
• Subject: Atmosphere / Atmospheric models / Climate / Climate change / Climate models / Climatology / Cloud formation / Clouds / Convection / Datasets / General circulation models / Geography / hybrid simulation / Machine learning / Microphysics / Neural networks / Parameterization / Physics / Seasonal variations / Seasonality / Simulation / Water vapor / Water vapour