Explore the vast range of titles available.

Global climate models are central tools for understanding past and future climate change. The assessment of model skill, in turn, can benefit from modern data science approaches. Here we apply causal discovery algorithms to sea level pressure data from a large set of climate model simulations and, as a proxy for observations, meteorological reanalyses. We demonstrate how the resulting causal networks (fingerprints) offer an objective pathway for process-oriented model evaluation. Models with fingerprints closer to observations better reproduce important precipitation patterns over highly populated areas such as the Indian subcontinent, Africa, East Asia, Europe and North America. We further identify expected model interdependencies due to shared development backgrounds. Finally, our network metrics provide stronger relationships for constraining precipitation projections under climate change as compared to traditional evaluation metrics for storm tracks or precipitation itself. Such emergent relationships highlight the potential of causal networks to constrain longstanding uncertainties in climate change projections. Algorithms to assess causal relationships in data sets have seen increasing applications in climate science in recent years. Here, the authors show that these techniques can help to systematically evaluate the performance of climate models and, as a result, to constrain uncertainties in future climate change projections.

Analysis of observations and model projections provides large-scale emergent constraints on the extent of CO 2 fertilization, with estimated increases in gross primary productivity for both high-latitude and extratropical ecosystems under elevated atmospheric CO 2 concentrations. Atmospheric CO 2 and plant growth Climate–carbon cycle models generally assume that elevated atmospheric CO 2 concentrations will enhance terrestrial plant productivity. But the magnitude of this so-called CO 2 fertilization effect remains uncertain—with implications for future climate change projections. This study provides large-scale constraints on the extent of CO 2 fertilization with an estimated increase in gross primary productivity of 37 ± 9 per cent for high-latitude ecosystems and 32 ± 9 per cent for extratropical ecosystems for a doubling of atmospheric CO 2 concentrations. Uncertainties in the response of vegetation to rising atmospheric CO 2 concentrations 1 , 2 contribute to the large spread in projections of future climate change 3 , 4 . Climate–carbon cycle models generally agree that elevated atmospheric CO 2 concentrations will enhance terrestrial gross primary productivity (GPP). However, the magnitude of this CO 2 fertilization effect varies from a 20 per cent to a 60 per cent increase in GPP for a doubling of atmospheric CO 2 concentrations in model studies 5 , 6 , 7 . Here we demonstrate emergent constraints 8 , 9 , 10 , 11 on large-scale CO 2 fertilization using observed changes in the amplitude of the atmospheric CO 2 seasonal cycle that are thought to be the result of increasing terrestrial GPP 12 , 13 , 14 . Our comparison of atmospheric CO 2 measurements from Point Barrow in Alaska and Cape Kumukahi in Hawaii with historical simulations of the latest climate–carbon cycle models demonstrates that the increase in the amplitude of the CO 2 seasonal cycle at both measurement sites is consistent with increasing annual mean GPP, driven in part by climate warming, but with differences in CO 2 fertilization controlling the spread among the model trends. As a result, the relationship between the amplitude of the CO 2 seasonal cycle and the magnitude of CO 2 fertilization of GPP is almost linear across the entire ensemble of models. When combined with the observed trends in the seasonal CO 2 amplitude, these relationships lead to consistent emergent constraints on the CO 2 fertilization of GPP. Overall, we estimate a GPP increase of 37 ± 9 per cent for high-latitude ecosystems and 32 ± 9 per cent for extratropical ecosystems under a doubling of atmospheric CO 2 concentrations on the basis of the Point Barrow and Cape Kumukahi records, respectively.

A promising method for improving the representation of clouds in climate models, and hence climate projections, is to develop machine learning‐based parameterizations using output from global storm‐resolving models. While neural networks (NNs) can achieve state‐of‐the‐art performance within their training distribution, they can make unreliable predictions outside of it. Additionally, they often require post‐hoc tools for interpretation. To avoid these limitations, we combine symbolic regression, sequential feature selection, and physical constraints in a hierarchical modeling framework. This framework allows us to discover new equations diagnosing cloud cover from coarse‐grained variables of global storm‐resolving model simulations. These analytical equations are interpretable by construction and easily transferable to other grids or climate models. Our best equation balances performance and complexity, achieving a performance comparable to that of NNs (R2 = 0.94) while remaining simple (with only 11 trainable parameters). It reproduces cloud cover distributions more accurately than the Xu‐Randall scheme across all cloud regimes (Hellinger distances < 0.09), and matches NNs in condensate‐rich regimes. When applied and fine‐tuned to the ERA5 reanalysis, the equation exhibits superior transferability to new data compared to all other optimal cloud cover schemes. Our findings demonstrate the effectiveness of symbolic regression in discovering interpretable, physically‐consistent, and nonlinear equations to parameterize cloud cover. Plain Language Summary In climate models, cloud cover is usually expressed as a function of coarse, pixelated variables. Traditionally, this functional relationship is derived from physical assumptions. In contrast, machine learning (ML) approaches, such as neural networks, sacrifice interpretability for performance. In our approach, we use high‐resolution climate model output to learn a hierarchy of cloud cover schemes from data. To bridge the gap between simple statistical methods and ML algorithms, we employ a symbolic regression method. Unlike classical regression, which requires providing a set of basis functions from which the equation is composed of, symbolic regression only requires mathematical operators (such as +, ×) that it learns to combine. By using a genetic algorithm, inspired by the process of natural selection, we discover an interpretable, nonlinear equation for cloud cover. This equation is simple, performs well, satisfies physical principles, and outperforms other algorithms when applied to new observationally‐informed data. Key Points We systematically derive and evaluate cloud cover parameterizations of various complexity from global storm‐resolving simulation output Using symbolic regression combined with physical constraints, we find a new interpretable equation balancing performance and simplicity Our data‐driven cloud cover equation can be retuned with few samples, facilitating transfer learning to generalize to other realistic data

Projections of future climate change play a fundamental role in improving understanding of the climate system as well as characterizing societal risks and response options. The Scenario Model Intercomparison Project (ScenarioMIP) is the primary activity within Phase 6 of the Coupled Model Intercomparison Project (CMIP6) that will provide multi-model climate projections based on alternative scenarios of future emissions and land use changes produced with integrated assessment models. In this paper, we describe ScenarioMIP's objectives, experimental design, and its relation to other activities within CMIP6. The ScenarioMIP design is one component of a larger scenario process that aims to facilitate a wide range of integrated studies across the climate science, integrated assessment modeling, and impacts, adaptation, and vulnerability communities, and will form an important part of the evidence base in the forthcoming Intergovernmental Panel on Climate Change (IPCC) assessments. At the same time, it will provide the basis for investigating a number of targeted science and policy questions that are especially relevant to scenario-based analysis, including the role of specific forcings such as land use and aerosols, the effect of a peak and decline in forcing, the consequences of scenarios that limit warming to below 2°C, the relative contributions to uncertainty from scenarios, climate models, and internal variability, and long-term climate system outcomes beyond the 21st century. To serve this wide range of scientific communities and address these questions, a design has been identified consisting of eight alternative 21st century scenarios plus one large initial condition ensemble and a set of long-term extensions, divided into two tiers defined by relative priority. Some of these scenarios will also provide a basis for variants planned to be run in other CMIP6-Endorsed MIPs to investigate questions related to specific forcings. Harmonized, spatially explicit emissions and land use scenarios generated with integrated assessment models will be provided to participating climate modeling groups by late 2016, with the climate model simulations run within the 2017-2018 time frame, and output from the climate model projections made available and analyses performed over the 2018-2020 period.

To mark the tenth anniversary of Nature Climate Change, we asked a selection of researchers across the broad range of climate change disciplines to share their thoughts on notable developments of the past decade, as well as their hopes and expectations for the coming years of discovery.

Cloud-related parameterizations remain a leading source of uncertainty in climate projections. Although machine learning holds promise for Earth system models (ESMs), many data-driven parameterizations lack interpretability, physical consistency, and smooth integration into ESMs. Here, a two-step method is presented to improve a climate model with data-driven parameterizations. First, we incorporate a physically consistent cloud cover parameterization—derived from storm-resolving simulations via symbolic regression, preserving interpretability while enhancing accuracy—into the ICON global atmospheric model. Second, we apply the gradient-free Nelder–Mead optimizer to automatically recalibrate the hybrid model against Earth observations, tuning in nested stages (2-, 7-, 30- and 365-day runs) to ensure stability and tractability. The tuned hybrid model substantially reduces long-standing biases in cloud cover—particularly over the Southern Ocean (by 75%) and subtropical stratocumulus regions (by 44%)—and remains robust under +4K surface warming. These results demonstrate that interpretable machine-learned parameterizations, paired with practical tuning, can efficiently and transparently strengthen ESM fidelity.

The Aerosol Chemistry Model Intercomparison Project (AerChemMIP) is endorsed by the Coupled-Model Intercomparison Project 6 (CMIP6) and is designed to quantify the climate and air quality impacts of aerosols and chemically reactive gases. These are specifically near-term climate forcers (NTCFs: methane, tropospheric ozone and aerosols, and their precursors), nitrous oxide and ozone-depleting halocarbons. The aim of AerChemMIP is to answer four scientific questions. 1. How have anthropogenic emissions contributed to global radiative forcing and affected regional climate over the historical period? 2. How might future policies (on climate, air quality and land use) affect the abundances of NTCFs and their climate impacts? 3.How do uncertainties in historical NTCF emissions affect radiative forcing estimates? 4. How important are climate feedbacks to natural NTCF emissions, atmospheric composition, and radiative effects? These questions will be addressed through targeted simulations with CMIP6 climate models that include an interactive representation of tropospheric aerosols and atmospheric chemistry. These simulations build on the CMIP6 Diagnostic, Evaluation and Characterization of Klima (DECK) experiments, the CMIP6 historical simulations, and future projections performed elsewhere in CMIP6, allowing the contributions from aerosols and/or chemistry to be quantified. Specific diagnostics are requested as part of the CMIP6 data request to highlight the chemical composition of the atmosphere, to evaluate the performance of the models, and to understand differences in behaviour between them.

Machine learning (ML)‐based parameterizations have been developed for Earth System Models (ESMs) with the goal to better represent subgrid‐scale processes or to accelerate computations. ML‐based parameterizations within hybrid ESMs have successfully learned subgrid‐scale processes from short high‐resolution simulations. However, most studies used a particular ML method to parameterize the subgrid tendencies or fluxes originating from the compound effect of various small‐scale processes (e.g., radiation, convection, gravity waves) in mostly idealized settings or from superparameterizations. Here, we use a filtering technique to explicitly separate convection from these processes in simulations with the Icosahedral Non‐hydrostatic modeling framework (ICON) in a realistic setting and benchmark various ML algorithms against each other offline. We discover that an unablated U‐Net, while showing the best offline performance, learns reverse causal relations between convective precipitation and subgrid fluxes. While we were able to connect the learned relations of the U‐Net to physical processes this was not possible for the non‐deep learning‐based Gradient Boosted Trees. The ML algorithms are then coupled online to the host ICON model. Our best online performing model, an ablated U‐Net excluding precipitating tracer species, indicates higher agreement for simulated precipitation extremes and mean with the high‐resolution simulation compared to the traditional scheme. However, a smoothing bias is introduced both in water vapor path and mean precipitation. Online, the ablated U‐Net significantly improves stability compared to the non‐ablated U‐Net and runs stable for the full simulation period of 180 days. Our results hint to the potential to significantly reduce systematic errors with hybrid ESMs. Plain Language Summary Due to their computational costs, it is currently not feasible to run more accurate high‐resolution climate models on a global domain on climate (century) time‐scales. However, high‐accuracy climate simulations are needed for more robust and detailed projections of our future climate. Here, we develop and evaluate various machine learning‐based convection parameterizations learned on reconstructed and coarse‐grained high‐resolution subgrid fluxes to solve this problem, and benchmark their performance. The data set is chosen from simulations of the Icosahedral Non‐hydrostatic modeling framework (ICON) in a realistic setting of the tropical Atlantic and at storm‐resolving resolutions. We focus only on convective subgrid fluxes that are isolated from other components. We improve the best ML algorithms further by excluding variables that cause unphysical correlations. Finally, we explain the learned relations of the best data‐driven schemes based on physical process understanding, test their performance when coupled to the ICON model, and achieve stable coupled simulations for 180 days as well as improved precipitation predictions. Key Points We train/benchmark machine learning models on convective fluxes derived from realistic coarse‐grained data of storm‐resolving simulations Shapley values reveal that the best offline model, a U‐Net, learns non‐causal links to precipitation and shows poor online performance A model, without non‐causal precipitation connections, runs more stable coupled to ICON and indicates better precipitation predictions

This paper describes the second major release of the Earth System Model Evaluation Tool (ESMValTool), a community diagnostic and performance metrics tool for the evaluation of Earth system models (ESMs) participating in the Coupled Model Intercomparison Project (CMIP). Compared to version 1.0, released in 2016, ESMValTool version 2.0 (v2.0) features a brand new design, with an improved interface and a revised preprocessor. It also features a significantly enhanced diagnostic part that is described in three companion papers. The new version of ESMValTool has been specifically developed to target the increased data volume of CMIP Phase 6 (CMIP6) and the related challenges posed by the analysis and the evaluation of output from multiple high-resolution or complex ESMs. The new version takes advantage of state-of-the-art computational libraries and methods to deploy an efficient and user-friendly data processing. Common operations on the input data (such as regridding or computation of multi-model statistics) are centralized in a highly optimized preprocessor, which allows applying a series of preprocessing functions before diagnostics scripts are applied for in-depth scientific analysis of the model output. Performance tests conducted on a set of standard diagnostics show that the new version is faster than its predecessor by about a factor of 3. The performance can be further improved, up to a factor of more than 30, when the newly introduced task-based parallelization options are used, which enable the efficient exploitation of much larger computing infrastructures. ESMValTool v2.0 also includes a revised and simplified installation procedure, the setting of user-configurable options based on modern language formats, and high code quality standards following the best practices for software development.

Persistent systematic errors in Earth system models (ESMs) arise from difficulties in representing the full diversity of subgrid, multiscale atmospheric convection and turbulence. Machine learning (ML) parameterizations trained on short high‐resolution simulations show strong potential to reduce these errors. However, stable long‐term atmospheric simulations with hybrid (physics + ML) ESMs remain difficult, as neural networks (NNs) trained offline often destabilize online runs. Training convection parameterizations directly on coarse‐grained data is challenging, notably because scales cannot be cleanly separated. This issue is mitigated using data from superparameterized simulations, which provide clearer scale separation. Yet, transferring a parameterization from one ESM to another remains difficult due to distribution shifts that induce large inference errors. Here, we present a proof‐of‐concept where a ClimSim‐trained, physics‐informed NN convection parameterization is successfully transferred to ICON‐A. The scheme is (a) trained on adjusted ClimSim data with subtracted radiative tendencies, and (b) integrated into ICON‐A. The NN parameterization predicts its own error, enabling mixing with a conventional convection scheme when confidence is low, thus making the hybrid AI‐physics model tunable with respect to observations and reanalysis through mixing parameters. This improves process understanding by constraining convective tendencies across column water vapor, lower‐tropospheric stability, and geographical conditions, yielding interpretable regime behavior. In Atmospheric Model Intercomparison Project‐style setups, several hybrid configurations outperform the default convection scheme (e.g., improved precipitation statistics). With additive input noise during training, both hybrid and pure‐ML schemes lead to stable simulations and remain physically consistent for at least 20 years, demonstrating inter‐ESM transferability and advancing long‐term integrability. Clouds and thunderstorms are difficult to simulate accurately in climate models because they typically occur at scales smaller than the model's grid. This necessitates the use of approximations for these processes, so‐called parameterizations, which often introduce errors. Machine learning (ML) offers a new way to improve these models, but ML can be unstable and doesn't always behave well when employed in different models or with different conditions. In this study, we develop a new hybrid method that combines machine learning with established physical principles to better simulate the influence of atmospheric convection. Our approach learns from high‐fidelity climate simulations and can adjust its behavior based on how confident the ML model is in its predictions. This helps the model stay stable and accurate, even when it is used in a different climate model. Furthermore, a small amount of noise is added during training to improve the long‐term stability of our ML model. We tested our method in the ICON climate model and found that it is accurate and stable in year‐long simulations, while remaining stable and reliable over periods of 20 years. This work shows that blending physics with machine learning can lead to more accurate and robust climate models. An machine learning convection parameterization trained on ClimSim and coupled to ICON achieves stable and accurate 20‐year Atmospheric Model Intercomparison Project simulations Physics‐informed loss, confidence‐guided mixing, and noise‐augmented training enhance conservation, accuracy, and stability, respectively The scheme can be tuned with observations by mixing in the conventional scheme when neural network confidence is low in moist, unstable regimes