This study explores integrating reinforcement learning (RL) with idealised climate models to address key parameterisation challenges in climate science. Current climate models rely on complex mathematical parameterisations to represent sub-grid scale processes, which can introduce substantial uncertainties. RL offers capabilities to enhance these parameterisation schemes, including direct interaction, handling sparse or delayed feedback, continuous online learning, and long-term optimisation. We evaluate the performance of eight RL algorithms on two idealised environments: one for temperature bias correction, another for radiative-convective equilibrium (RCE) imitating real-world computational constraints. Results show different RL approaches excel in different climate scenarios with exploration algorithms performing better in bias correction, while exploitation algorithms proving more effective for RCE. These findings support the potential of RL-based parameterisation schemes to be integrated into global climate models, improving accuracy and efficiency in capturing complex climate dynamics. Overall, this work represents an important first step towards leveraging RL to enhance climate model accuracy, critical for improving climate understanding and predictions. Code accessible at https://github.com/p3jitnath/climate-rl.

Tropical high cloud cover decreases with surface warming in most general circulation models. This reduction, according to the “stability-iris” hypothesis, is thermodynamically controlled and linked to a decrease in the radiatively-driven clear-sky convergence, when the peak anvil clouds rise because of the rising isotherms. The influence of the large-scale dynamical changes on the tropical high cloud fraction remains difficult to disentangle from the local thermodynamic influence, given that the mean meridional circulation remains inextricably tied to the local thermodynamic structure of the atmosphere. However, using idealized general circulation model (GCM) simulations, we propose a novel method to segregate the dynamical impact from the thermodynamic impact on the tropical high cloud fraction. To this end, our investigation primarily focuses on the mechanisms underpinning changes in the high cloud cover in the deep tropics in response to extratropical surface warming, when tropical sea surface temperatures remain invariant. We find that the relative importance of the net convective detrainment of ice cloud condensates to the cloud microphysical processes, such as the net depositional growth of ice aggregates, in controlling the tropical high cloud fraction is altitude-dependent.

We investigate positive subtropical low cloud feedback mechanisms in climate models which have performed the CMIP6/CFMIP-3 AMIP and AMIP uniform +4K experiments while saving CFMIP-3 process diagnostics on model levels. Our analysis focuses on the trade cumulus/stratocumulus transition region between California and Hawaii, where positive low cloud feedbacks are present in the JJA season. We introduce a methodology to} test various positive cloud feedback mechanisms proposed in the literature as primary explanations for the low cloud responses in the models. Causal hypotheses are tested by comparing their predictions with the models’ responses of clouds, cloud controlling factors, boundary layer depth and temperature/humidity tendencies to climate warming. Changes in boundary layer depth, relative humidity in the cloud layer and humidity advection at the top of the boundary layer are shown to distinguish among the hypotheses considered. For the cases examined, our approach rules out 4/5 of the mechanisms considered in half of the models and 3/5 in the remainder. We argue that unambiguously identifying the positive feedback mechanisms operating in models will in some cases require intervention experiments designed to test specific hypotheses.

Climate feedbacks over the historical period (1850–2014) have been investigated in large ensembles of historical, hist-ghg, hist-aer, and hist-nat experiments, with 47 members for each experiment. Across the historical ensemble with all forcings, a range in estimated Effective Climate Sensitivity (EffCS) between approximately 3–6 K is found, a considerable spread stemming solely from initial condition uncertainty. The spread in EffCS is associated with varying Sea Surface Temperature (SST) patterns seen across the ensemble due to their influence on different feedback processes. For example, the level of polar amplification is shown to strongly control the amount of sea ice melt per degree of global warming. This mechanism is responsible for the large spread in shortwave clear-sky feedbacks and is the main contributor to the different forcing efficacies seen across the different forcing agents, although in HadGEM3-GC3.1-LL these differences in forcing efficacy are shown to be small. The spread in other feedbacks is also investigated, with the level of tropical SST warming shown to strongly control the longwave clear-sky feedbacks, and the local surface-air-temperatures and large scale tropospheric temperatures shown to influence cloud feedbacks. The metrics used to understand the spread in feedbacks can also help to explain the disparity between feedbacks seen in the historical experiment simulations and a more accurate modeled estimate of the feedbacks seen in the real world derived from an atmosphere-only experiment prescribed with observed SSTs (termed amip-piForcing).

Low cloud feedback in global warming projections by climate models is characterized by its positive sign, the mechanism of which is not well understood. Here we propose that the positive sign is primarily caused by the increase in upward longwave radiation from the sea surface. We devise numerical experiments that enable separation of the feedback into components coming from physically distinct causes. Results of these experiments with a climate model indicate that increases in upward longwave radiation from the sea surface cause warming and absolute drying in the boundary layer, leading to the positive low cloud feedback. The absolute drying results from decrease in surface evaporation, and also from decrease in inversion strength which enhances vertical mixing of drier free tropospheric air into the boundary layer. This mechanism is different from previously proposed understanding that positive low cloud feedback is caused by increases in surface evaporation or vertical moisture contrast.

We compare the radiative feedbacks resulting from a uniform warming and cooling of sea surface temperatures by 4 K in an ensemble of global climate models. The global-mean net feedback is less stabilising in response to warming in all nine models. This is primarily due to a stronger tropical water vapour feedback, with a smaller contribution from the shortwave cloud feedback. The zonal-mean feedbacks are similarly robust across the ensemble. In the extra-tropics, more positive shortwave cloud feedback under warming is associated with further poleward migration of the mean Southern Hemisphere jet latitude in some models. However, additional experiments with an aquaplanet version of the HadGEM3 model suggest that the asymmetry of the jet shift is not driving that in the cloud feedbacks at these latitudes. In the tropics, stronger water vapour feedback under warming is offset by a weaker shortwave cloud feedback. The result is that the ensemble spread in the differences between the global feedbacks under warming and cooling is mainly determined by their differences in the tropics. The spatial distribution of the feedbacks largely reflects the zonal mean behaviour, although there is considerable intermodel variation in the regional cloud feedbacks, particularly in the tropical shortwave cloud feedback. Comparison with CO2- and solar-forced coupled experiments suggests that the global-mean longwave cloud feedback is nearly invariant to warming and cooling, regardless of the nature of the forcing. The shortwave cloud feedback is generally more positive under warming in the coupled models, consistent with the uniform SST perturbation experiments.

Continuous Structural Parameterization (CSP) is a method for approximating different numerical model parameterizations of the same process as functions of the same gridscale variables. This allows systematic comparison of parameterizations with each other and observations or resolved simulations of the same process. Using the example of two convection schemes running in the Met Office Unified Model (UM), we show that a CSP is able to capture concisely the broad behavior of the two schemes, and differences between the parameterizations and resolved convection simulated by a high resolution simulation. When the original convection schemes are replaced with their CSP emulators within the UM, basic features of the original model climate and some features of climate change are reproduced, demonstrating that CSP can capture much of the important behavior of the schemes. Our results open the possibility that future work will estimate uncertainty in model projections of climate change from estimates of uncertainty in simulation of the relevant physical processes.