Mark. J. Webb

Weather and climate models rely on parameterizations to represent unresolved sub‐grid processes. Traditional schemes rely on fixed coefficients that are weakly constrained and tuned offline, contributing to persistent biases that limit their ability to adapt to underlying physics. This study presents a framework that learns components of parametrization schemes online as a function of the evolving model state using reinforcement learning (RL) and evaluates policy‐driven parameter updates across idealized testbeds spanning a simple climate bias correction (SCBC), a radiative‐convective equilibrium (RCE), and a zonal mean energy balance model (EBM) with single‐agent and federated multi‐agent settings. Across nine RL algorithms, Truncated Quantile Critics (TQC), Deep Deterministic Policy Gradient (DDPG), and Twin Delayed DDPG (TD3) achieved the highest skill and stable convergence, with performance assessed against a static baseline using area‐weighted RMSE, temperature and pressure‐level diagnostics. For the EBM, single‐agent RL outperformed static parameter tuning with the strongest gains in tropical and mid‐latitude bands, while federated RL on multi‐agent setups enabled specialized control and faster convergence, with a six‐agent DDPG configuration using frequent aggregation yielding the lowest area‐weighted RMSE across the tropics and mid‐latitudes. The learned corrections were also physically meaningful as agents modulated EBM radiative parameters to reduce meridional biases, adjusted RCE lapse rates to match vertical temperature errors, and stabilized heating increments to limit drift. Overall, results show that RL can learn skillful state‐dependent parametrization components in idealized settings, offering a scalable pathway for online learning within numerical models and a starting point for evaluation in weather and climate models.

Weather and climate models rely on parametrisations to represent unresolved sub-grid processes. Traditional schemes rely on fixed coefficients that are weakly constrained and tuned offline, contributing to persistent biases that limit their ability to adapt to underlying physics. This study presents a framework that learns components of parametrisation schemes online as a function of the evolving model state using reinforcement learning (RL) and evaluates policy-driven parameter updates across idealised testbeds spanning a simple climate bias correction (SCBC), a radiative-convective equilibrium (RCE), and a zonal mean energy balance model (EBM) with single-agent and federated multi-agent settings. Across nine RL algorithms, Truncated Quantile Critics (TQC), Deep Deterministic Policy Gradient (DDPG), and Twin Delayed DDPG (TD3) achieved the highest skill and stable convergence, with performance assessed against a static baseline using area-weighted RMSE, temperature and pressure-level diagnostics. For the EBM, single-agent RL outperformed static parameter tuning with the strongest gains in tropical and mid-latitude bands, while federated RL on multi-agent setups enabled specialised control and faster convergence, with a six-agent DDPG configuration using frequent aggregation yielding the lowest area-weighted RMSE across the tropics and mid-latitudes. The learnt corrections were also physically meaningful as agents modulated EBM radiative parameters to reduce meridional biases, adjusted RCE lapse rates to match vertical temperature errors, and stabilised heating increments to limit drift. Overall, results show that RL can learn skilful state-dependent parametrisation components in idealised settings, offering a scalable pathway for online learning within numerical models and a starting point for evaluation in weather and climate models.

Low clouds (LC) are the largest source of uncertainty in the climate response to increased in global climate models. This is due to the multitude of factors that affect their behavior and how these are characterized across models. Previous work has explored how the deep convective tropical ascent regions can affect LC in subsidence regimes. In this study we apply a cooling in the free troposphere over the tropical West Pacific (WP) to induce a weakening in deep convection in an atmosphere‐only configuration of the HadGEM3 model. This weakened WP ascent causes a reduction in LC in the East Pacific (EP) subsiding region. Our hypothesized mechanism is that the weakened ascent leads to a weakening of lower tropospheric advection across the equatorial Pacific, decreasing moistening in the lower free troposphere. This reduction in moisture increases net longwave cooling immediately above the boundary layer, which increases cloud‐top cooling and cloud‐top entrainment of dry air, deepening the boundary layer while drying the air below, leading to a reduction in LC fraction. Finally, this causes a positive local cloud radiative effect response, the same sign as that expected from predicted greenhouse gas driven climate change. Although the inversion strength plays a large role in LC reduction in some parts of the EP, it is not the main cause at all locations and still does not account for the entirety of LC reduction. Other potential mechanisms, such as changes in subsidence, are explored and ruled out as the main cause for the change in LC.

The Sherwood et al. assessment [1] of Earth's climate sensitivity to a doubling of atmospheric carbon dioxide concentration broke new ground in providing estimates of radiative feedback and its components through the use of multiple lines of evidence. The assessment combined evidence from Global Climate Models (GCMs) with evidence from observations and process models that are able to produce more defensible estimates of small-scale and poorly-understood processes. However, by treating estimates of the different components of feedbacks as independent of one another, Sherwood ignored correlations between different feedbacks, which could impact the uncertainty affecting the estimate of overall feedback. The exception to this was the well-known water vapour-lapse rate anti-correlation, which they did consider.In this study, we first undertake a perfect model experiment with the CMIP5 and CMIP6 ensembles that demonstrates the effects of considering correlations between components of feedbacks on estimates of net radiative feedback in a Sherwood-type analysis. Second, we explore correlations between contemporary estimates of feedback components from observed climate variability and cloud controlling factor analysis. Correlations between components have a similar structure for both perfect model and contemporary estimates. It is found that introducing feedback correlations into the Sherwood framework increases the standard deviation of the net feedback uncertainty by about 30 %. Impacts on estimates of climate sensitivity are smaller, because the process-based estimate of radiative feedback is only one part of the sensitivity estimate.Prospects for future feedback and sensitivity estimates are discussed. The caveat to our results is that Sherwood's estimates of feedback components come from different sources. Although our results suggest that at least some of these show similar correlation structures, there is a need for future work that aims to understand the physical and statistical relationships between estimates of different components of feedback.Reference[1] Sherwood et al., 2020, Rev. Geophys., https://doi.org/10.1029/2019RG000678.

Sub-grid parameterisations in climate models are traditionally static and tuned offline, limiting adaptability to evolving states. This work introduces FedRAIN-Lite , a federated reinforcement learning (FedRL) framework that mirrors the spatial decomposition used in general circulation models (GCMs) by assigning agents to latitude bands, enabling local parameter learning with periodic global aggregation. Using a hierarchy of simplified energy-balance climate models, from a single-agent baseline ( ebm-v1 ) to multi-agent ensemble ( ebm-v2 ) and GCM-like ( ebm-v3 ) setups, we benchmark three RL algorithms under different FedRL configurations. Results show that Deep Deterministic Policy Gradient (DDPG) consistently outperforms both static and single-agent baselines, with faster convergence and lower area-weighted RMSE in tropical and mid-latitude zones across both ebm-v2 and ebm-v3 setups. DDPG’s ability to transfer across hyperparameters and low computational cost make it well suited for geographically adaptive parameter learning. This capability offers a scalable pathway towards high-complexity GCMs and provides a prototype for physically aligned, online-learning climate models that can evolve with a changing climate. Code accessible at https://github.com/p3jitnath/climate-rl-fedrl .

HadGEM3-GC3.1-LL exhibits strong positive shortwave cloud feedbacks in the subtropical stratocumulus/trade cumulus transition regions, which contribute to its high climate sensitivity. We investigate the causes of low-level cloud decreases in this regime by separating the effects of increasing SST on upwelling longwave radiation from the surface (SST-rad) from SST-driven effects on turbulent surface sensible and latent heat fluxes (SST-turb) in HadGEM3-GC3.1-LL, following the approach of Ogura et al. (2023). The positive low cloud feedback in the transition regime between California and Hawaii in HadGEM3-GC3.1-LL is mainly caused by the SST-driven increase in upwelling longwave radiation from the surface, as was shown previously for MIROC5 and MIROC6 in Ogura et al. (2023). SST-driven changes in sensible and latent heat fluxes make a smaller contribution to the positive feedback in this region. Repeating these HadGEM3-GC3.1-LL experiments with the convective parametrization deactivated shows that parametrized convection amplifies the low cloud feedback by changing the sign of the SST-turb feedback component from negative to positive. This is due to a muted cloud fraction increase near cloud base when parametrized convection is active, consistent with weakening surface wind speeds, surface buoyancy fluxes and increasing critical relative humidity, or with stabilisation of the boundary layer causing increased heating of the cloud layer by compensating subsidence.

Subtropical low clouds (LC) are the largest source of uncertainty in the climate response to increased CO2 in global climate models (GCMs). This is due to the multitude of factors that affect their behaviour and how these are characterized across models. Previous work has explored how the deep convective tropical ascent regions can affect subtropical LC. In this study we apply a cooling in the free troposphere over the tropical West Pacific to induce a weakening in deep convection in an atmosphere-only configuration of the HadGEM3 model. This weakened ascent causes a reduction in LC in the subtropical East Pacific. Our hypothesized mechanism is that the weakened ascent leads to a weakening of lower tropospheric advection across the equatorial Pacific, decreasing moistening in the lower free troposphere. This reduction in moisture increases net long wave cooling immediately above the boundary layer, which increases cloud-top cooling and cloud-top entrainment of dry air, deepening the boundary layer while drying the air below, leading to a reduction in low cloud fraction. Finally, this causes a positive local cloud radiative effect response, the same sign as that expected from predicted greenhouse gas driven climate change. Other potential mechanisms, such as changes in subtropical subsidence, are explored and ruled out as the cause for the change in LC. We demonstrate how regional forcing can induce global change, highlighting the necessity of understanding regional contributions to predicted future climate scenarios.

Centennial oscillations occur in the Southern Ocean in the majority of an ensemble of pre-industrial control runs of 16 CMIP6 models. These oscillations involve features such as the Antarctic Circumpolar Current, deep convection, and the lower meridional circulation cell. The oscillations persist throughout historical forcing runs but cease in the following 30-50 years (with the range of oscillation decreasing from 15 Sv to 5 Sv in the CanESM5 model), depending on climate forcing. Alongside this, the lower cell also shows a large reduction of 21+-17% under SSP585 forcing across the ensemble, and more than 50% in some models, with a net reduction under all but one of the scenarios investigated. This suggests that historical forcing already experienced is sufficient to cross a ”tipping point” and shut down the oscillations and lower cell, even under strong climate change mitigation. These oscillations dominate the model internal variability when they are present and so strongly influence the initial conditions for future climate projections and the timescale for lower cell shutdown, though not its final strength. They therefore dominate our predictions for changes in Southern Ocean features such as the ACC, increasing uncertainty and masking climate forcing signals. The dynamical response of the oscillations to warming and the impact of initial oscillation phase on this response are investigated through idealised forcing runs of the HadGEM3-GC3.1-LL model. These analyses suggest that Southern Ocean oscillations, and their response to climate forcing, are dominated by changes in deep convection around Antarctica and dense water formation and export.