Machine-learning (ML) parameterizations of subgrid processes (here of turbulence, convection, and radiation) may one day replace conventional parameterizations by emulating high-resolution physics without the cost of explicit simulation. However, their development has been stymied by uncertainty surrounding whether or not improved offline performance translates to improved online performance (i.e., when coupled to a large-scale general circulation model (GCM)). A key barrier has been the limited sampling of the online effects of the ML design decisions and tuning due to the complexity of performing large ensembles of hybrid physics-ML climate simulations. Our work examines the coupled behavior of full-physics ML parameterizations using large ensembles of hybrid simulations, totalling 2,970 in our case. With extensive sampling, we statistically confirm that lowering offline error lowers online error (given certain constraints). However, we also reveal that decisions decreasing online error, like removing dropout, can trade off against hybrid model stability and vice versa. Nevertheless, we are able to identify design decisions that yield unambiguous improvements to offline and online performance, namely incorporating memory and training on multiple climates. We also find that converting moisture input from specific to relative humidity enhances online stability and that using a Mean Absolute Error (MAE) loss breaks the aforementioned offline/online error relationship. By enabling rapid online experimentation at scale, we empirically answer previously unresolved questions regarding subgrid ML parameterization design.

Can the current successes of global machine learning-based weather simulators be generalized beyond two-week forecasts to stable and accurate multiyear runs? The recently developed AI2 Climate Emulator (ACE) suggests this is feasible, based upon 10-year simulations with a network trained on output from a physics-based global atmosphere model using a grid spacing of approximately 110 km and forced by a repeating annual cycle of sea-surface temperature. Here we show that ACE, without modification, can be trained to emulate another major atmospheric model, EAMv2, run at a comparable grid spacing for at least ten years with similarly small climate biases -- a necessary prerequisite to wider applicability. With a comprehensive analysis from multiple temporal, spatial, and frequency domain perspectives, we show that ACE faithfully represents the spatiotemporal structure of EAMv2 precipitation and related variables. Finally, we show that a pretrained ACE network is able to adapt to a new global climate model simulation dataset with 10x fewer training steps than when starting from random initialization, all while still maintaining low levels of climate bias. Further analysis of these fine-tuning experiments reveal ACE's attractive ability to interpolate between distinct global climate models.

We investigate a set of Energy Exascale Earth System Model Multi-scale Modeling Framework (E3SM-MMF) simulations that vary the dimensionality and momentum transport configurations of the embedded cloud-resolving models (CRMs), including unusually ambitious 3D configurations. Issues endemic to all MMF simulations include too much ITCZ rainfall and too little over the Amazon. Systematic MMF improvements include more on-equatorial rainfall across the Warm Pool. Interesting sensitivities to CRM domain are found in the regional time-mean precipitation pattern over the tropics. The 2D E3SM-MMF produces an unrealistically rainy region over the northwestern tropical Pacific; this is reduced in computationally ambitious 3D configurations that use 1024 embedded CRM grid columns per host cell. Trajectory analysis indicates that these regional improvements are associated with desirably fewer tropical cyclones and less extreme precipitation rates. To understand why and how the representation of precipitation improved in 3D, we propose a framework that dilution is stronger in 3D. This viewpoint is supported by multiple indirect lines of evidence, including a delayed moisture-precipitation pickup, smaller precipitation efficiency, and amplified convective mass flux profiles and more high clouds. We also demonstrate that the effects of varying embedded CRM dimensionality and momentum transport on precipitation can be identified during the first few simulated days, providing an opportunity for rapid model tuning without high computational cost. Meanwhile the results imply that other less computationally intensive ways to enhance dilution within MMF CRMs may also be strategic tuning targets.

We design a new strategy to load-balance high-intensity sub-grid atmospheric physics calculations restricted to a small fraction of a global climate simulation’s domain. We show why the current parallel load balancing infrastructure of CESM and E3SM cannot efficiently handle this scenario at large core counts. As an example, we study an unusual configuration of the E3SM Multiscale Modeling Framework (MMF) that embeds a binary mixture of two separate cloud-resolving model grid structures that is attractive for low cloud feedback studies. Less than a third of the planet uses high-resolution (MMF-HR; sub-km horizontal grid spacing) relative to standard low-resolution (MMF-LR) cloud superparameterization elsewhere. To enable MMF runs with Multi-Domain CRMs, our load balancing theory predicts the most efficient computational scale as a function of the high-intensity work’s relative overhead and its fractional coverage. The scheme successfully maximizes model throughput and minimizes model cost relative to precursor infrastructure, effectively by devoting the vast majority of the processor pool to operate on the few high-intensity (and rate-limiting) HR grid columns. Two examples prove the concept, showing that minor artifacts can be introduced near the HR/LR CRM grid transition boundary on idealized aquaplanets, but are minimal in operationally relevant real-geography settings. As intended, within the high (low) resolution area, our Multi-Domain CRM simulations exhibit cloud fraction and shortwave reflection convergent to standard baseline tests that use globally homogenous MMF-LR and MMF-HR. We suggest this approach can open up a range of creative multi-resolution climate experiments without requiring unduly large allocations of computational resources.

The spread in global mean precipitation among climate models is explored in two ensembles using the complementary perspectives of surface evaporation and energy budgets. Models with higher global-mean precipitation have stronger oceanic evaporation, driven by drier near-surface air. The drier surface conditions occur alongside increases in near-surface temperature and moisture at 925 hPa, which point to stronger boundary layer mixing. Correlations suggest that the degree of lower tropospheric mixing explains 18%-49% of the intermodel precipitation variance. To test this hypothesis, the degree of mixing is varied in a single-model experiment by adjusting the relative humidity threshold that controls low-cloud fraction. Indeed, increasing lower tropospheric mixing results in more global mean precipitation. Energetically, increased precipitation rates are associated with more downwelling longwave radiation to the surface and weaker sensible heat fluxes. These results highlight how lower-tropospheric processes must be better constrained to reduce the precipitation discrepancy among climate models.

Key Points: • Machine learning (ML) helps model the interaction between clouds and climate using large datasets. • We review physics-guided/explainable ML applied to cloud-related processes in the climate system. • We also provide a guide to scientists who would like to get started with ML. Abstract: Machine learning (ML) algorithms are powerful tools to build models of clouds and climate that are more faithful to the rapidly-increasing volumes of Earth system data than commonly-used semiempirical models. Here, we review ML tools, including interpretable and physics-guided ML, and outline how they can be applied to cloud-related processes in the climate system, including radiation, microphysics, convection, and cloud detection , classification, emulation, and uncertainty quantification. We additionally provide a short guide to get started with ML and survey the frontiers of ML for clouds and climate .

Sub-kilometer processes are critical to the physics of aerosol-cloud interaction but have been dependent on parameterizations in global model simulations. We thus report the strength of aerosol-cloud interaction in the Ultra-Parameterized Community Atmosphere Model (UPCAM), a multiscale climate model that uses coarse exterior resolution to embed explicit cloud resolving models with enough resolution (250-m horizontal, 20-m vertical) to quasi-resolve sub-kilometer eddies. To investigate the impact on aerosol-cloud interactions, UPCAMâ\euro™s simulations are compared to a coarser multi-scale model with 3 km horizontal resolution. UPCAM produces cloud droplet number concentrations ($N_\mathrm{d}$) and cloud liquid water path (LWP) values that are higher than the coarser model but equally plausible compared to observations. Our analysis focuses on the Northern Hemisphere midlatitude oceans, where historical aerosol increases have been largest. We find similarities in the overall radiative forcing from aerosol-cloud interactions in the two models, but this belies fundamental underlying differences. The radiative forcing from increases in LWP is weaker in UPCAM, whereas the forcing from increases in $N_\mathrm{d}$ is larger. Surprisingly, the weaker LWP increase is not due to a weaker increase in LWP in raining clouds, but a combination of weaker increase in LWP in non-raining clouds and a smaller fraction of raining clouds in UPCAM. The implication is that as global modeling moves towards finer than storm-resolving grids, nuanced model validation of ACI statistics conditioned on the existence of precipitation and good observational constraints on the baseline probability of precipitation will become key for tighter constraints and better conceptual understanding.

Vertical resolution is an often overlooked parameter in simulations of convection. We explore the sensitivity of simulated deep convection to vertical resolution in the System for Atmospheric Modeling (SAM) convection resolving model. We analyze simulations run in tropical radiative convective equilibrium with 32, 64, 128, and 256 vertical levels in a small (100 km) and large domain (1500 km). At high vertical resolution, the relative humidity and anvil cloud fraction are reduced, which is linked to a reduction in both fractional and volumetric detrainment. This increases total atmospheric radiative cooling at high resolution, which leads to enhanced surface fluxes and precipitation, despite reduced column water vapor. In large domains, convective aggregation begins by simulation day 25 for simulations with 64 and 128 levels, while onset is delayed until simulation day 75 for the simulation with 32 vertical levels. Budget analyses reveal that mechanisms involved in the generation and maintenance of convective aggregation for the 32-level simulation differ from those for the 64- and 128-level simulations. Weaker cold pools in the 32-level simulation allow the boundary layer in dry regions to become extremely dry, which leads to an aggregated state with very strong spatial gradients in column-integrated moist static energy. Understanding both the triggering and maintenance of convective aggregation and its simulated sensitivity to model formulation is a necessary component of atmospheric modeling. We show that vertical resolution has a strong impact on the mean state and convective behavior in both small and large domains.