Traditional numerical global climate models simulate the full Earth system by exchanging boundary conditions between separate simulators of the atmosphere, ocean, sea ice, land surface, and other geophysical processes. This paradigm allows for distributed development of individual components within a common framework, unified by a coupler that handles translation between realms via spatial or temporal alignment and flux exchange. Following a similar approach adapted for machine learning-based emulators, we present SamudrACE: a coupled global climate model emulator which produces centuries-long simulations at 1-degree horizontal, 6-hourly atmospheric, and 5-daily oceanic resolution, with 145 2D fields spanning 8 atmospheric and 19 oceanic vertical levels, plus sea ice, surface, and top-of-atmosphere variables. SamudrACE is highly stable and has low climate biases comparable to those of its components with prescribed boundary forcing, with realistic variability in coupled climate phenomena such as ENSO that is not possible to simulate in uncoupled mode.

Modern climate projections lack adequate spatial and temporal resolution due to computational constraints, leading to inaccuracies in representing critical processes like thunderstorms that occur on the sub-resolution scale. Hybrid methods combining physics with machine learning (ML) offer faster, higher fidelity climate simulations by outsourcing compute-hungry, high-resolution simulations to ML emulators. However, these hybrid physics-ML simulations require domain-specific data and workflows that have been inaccessible to many ML experts. This paper is an extended version of our NeurIPS award-winning ClimSim dataset paper (Yu et al., 2024). The ClimSim dataset includes 5.7 billion pairs of multivariate input/output vectors spanning ten years at high temporal resolution, capturing the influence of high-resolution, high-fidelity physics on a host climate simulator's macro-scale state. In this extended version, we introduce a significant new contribution in Section 5, which provides a cross-platform, containerized pipeline to integrate ML models into operational climate simulators for hybrid testing. We also implement various baselines of ML models and hybrid simulators to highlight the ML challenges of building stable, skillful emulators.

Climate simulations, at all grid resolutions, rely on approximations that encapsulate the forcing due to unresolved processes on resolved variables, known as parameterizations. Parameterizations often lead to inaccuracies in climate models, with significant biases in the physics of key climate phenomena. Advances in artificial intelligence (AI) are now directly enabling the learning of unresolved processes from data to improve the physics of climate simulations. Here, we introduce a flexible framework for developing and implementing physics- and scale-aware machine learning parameterizations within climate models. We focus on the ocean and sea-ice components of a state-of-the-art climate model by implementing a spectrum of data-driven parameterizations, ranging from complex deep learning models to more interpretable equation-based models. Our results showcase the viability of AI-driven parameterizations in operational models, advancing the capabilities of a new generation of hybrid simulations, and include prototypes of fully coupled atmosphere-ocean-sea-ice hybrid simulations. The tools developed are open source, accessible, and available to all.

Herein, spatial variations of sea level trends from the altimeter record are compared to contemporaneous (1993-2014) and future trends in ocean dynamic sea level from state-of-the-art climate models. A multiclimate model ensemble of CMIP6 historical simulations is analyzed (n 5 560), and little agreement is found in the global pattern of ocean dynamic sea level trends across the ensemble. While some simulations have regional ocean dynamic sea level trends that are a close match to the altimeter record, none are a good match globally (maximum pattern correlation globally of 0.47 and 5%-95% range from 20.20 to 0.26), and simultaneously matching the altimeter record in the tropical and North Pacific and tropical and North Atlantic is particularly challenging. Our focus in this study is on differences across the individual historical simulations and the role for internal variability, external forcing, and structural factors in driving these differences. A close relationship is found between patterns of sea surface temperature trends and those in sea level, and both can be related to the trajectories of common modes of atmosphere-ocean variability, with centers of action in the Indian Ocean and the tropical and North Pacific. Using preindustrial control simulations, we determine where external forcing has and will produce local (i.e., gridpoint level) ocean dynamic sea level trends that are significant relative to internal variability. At the present (1992-2023), climate models suggest that ocean dynamic sea level trends over -15% of the ocean area are significant relative to internal variability, with this number increasing to 37% by 2050 under a high-emission scenario (33% under a low-emission scenario).

AI emulators for forecasting have emerged as powerful tools that can outperform conventional numerical predictions. The next frontier is to build emulators for long climate simulations with skill across a range of spatiotemporal scales, a particularly important goal for the ocean. Our work builds a skillful global emulator of the ocean component of a state-of-the-art climate model. We emulate key ocean variables, sea surface height, horizontal velocities, temperature, and salinity, across their full depth. We use a modified ConvNeXt UNet architecture trained on multi-depth levels of ocean data. We show that the ocean emulator-Samudra-which exhibits no drift relative to the truth, can reproduce the depth structure of ocean variables and their interannual variability. Samudra is stable for centuries and 150 times faster than the original ocean model. Samudra struggles to capture the correct magnitude of the forcing trends and simultaneously remain stable, requiring further work.

Accurate prediction of climate in the subseasonal-to-seasonal scale is crucial for disaster preparedness and robust decision making amidst climate change. Yet, forecasting beyond the weather timescale is challenging because it deals with problems other than initial condition, including boundary interaction, butterfly effect, and our inherent lack of physical understanding. At present, existing benchmarks tend to have shorter forecasting range of up-to 15 days, do not include a wide range of operational baselines, and lack physics-based constraints for explainability. Thus, we propose ChaosBench, a challenging benchmark to extend the predictability range of data-driven weather emulators to S2S timescale. First, ChaosBench is comprised of variables beyond the typical surface-atmospheric ERA5 to also include ocean, ice, and land reanalysis products that span over 45 years to allow for full Earth system emulation that respects boundary conditions. We also propose physics-based, in addition to deterministic and probabilistic metrics, to ensure a physically-consistent ensemble that accounts for butterfly effect. Furthermore, we evaluate on a diverse set of physics-based forecasts from four national weather agencies as baselines to our data-driven counterpart such as ViT/ClimaX, PanguWeather, GraphCast, and FourCastNetV2. Overall, we find methods originally developed for weather-scale applications fail on S2S task: their performance simply collapse to an unskilled climatology. Nonetheless, we outline and demonstrate several strategies that can extend the predictability range of existing weather emulators, including the use of ensembles, robust control of error propagation, and the use of physics-informed models. Our benchmark, datasets, and instructions are available at https://leap-stc.github.io/ChaosBench.

With the success of machine learning (ML) applied to climate reaching further every day, emulators have begun to show promise not only for weather but for multi-year time scales in the atmosphere. Similar work for the ocean remains nascent, with state-of-the-art limited to models running for shorter time scales or only for regions of the globe. In this work, we demonstrate high-skill global emulation for surface ocean fields over 5-8 years of model rollout, accurately representing modes of variability for two different ML architectures (ConvNext and Transformers). In addition, we address the outstanding question of generalization, an essential consideration if the end-use of emulation is to model warming scenarios outside of the model training data. We show that 1) generalization is not an intrinsic feature of a data-driven emulator, 2) fine-tuning the emulator on only small amounts of additional data from a distribution similar to the test set can enable the emulator to perform well in a warmed climate, and 3) the forced emulators are robust to noise in the forcing.

Understanding air-sea interaction is crucial for our ability to predict future states of the climate system, and to inform economic and societal decision-making. However, the representation of air-sea interactions in climate models is limited by structural errors associated with model resolution. Coarse-resolution climate models do not resolve small-scale structures in the air-sea state, which, due to strong nonlinearities in the coupling formulae, can impact the large-scale air-sea exchange—a mechanism that has received little attention and is the focus of this paper. Since observations at the temporal and spatial coverage needed to study this problem do not yet exist, we quantify the impact of this small-scale heterogeneity on the large-scale air-sea heat flux by analyzing 1/10° coupled climate simulations. This effect systematically cools the ocean by about 4W/m2 globally—with large spatio-temporal variations—and mostly enhances the large-scale heat flux. By identifying an overlooked contribution to air-sea heat flux in climate models, we open a promising new direction for addressing biases in climate simulations and thus improving future climate predictions. Furthermore, future observations, like the newly proposed satellite mission ODYSEA35, could potentially observe and quantify this effect directly.

Abstract When different flooding drivers co‐occur, they can cause compound floods. Despite the potential impact of compound flooding, few studies have projected how the joint probability of flooding drivers may change. Furthermore, existing projections may not be very robust, as they are based on only 5 to 6 climate model simulations. Here, we use a large ensemble of simulations from the Coupled Model Intercomparison Project 6 (CMIP6) to project changes in the joint probability of extreme storm surges and precipitation at European tide gauges under a medium and high emissions scenario, enabled by data‐proximate cloud computing and statistical storm surge modeling. We find that the joint probability will increase in the northwest and decrease in most of the southwest of Europe. Averaged over Europe, the absolute magnitude of these changes is 36%–49% by 2080, depending on the scenario. The large‐scale changes in the joint probability of extreme storm surges and precipitation are similar to those in the joint probability of extreme wind speeds and precipitation, but locally, differences can exceed the changes themselves. Due to internal climate variability and inter‐model differences, projections based on simulations of only 5 to 6 randomly chosen CMIP6 models have a probability of higher than 10% to differ qualitatively from projections based on all CMIP6 simulations in multiple regions, especially under the medium emissions scenario and earlier in the twenty‐first century. Therefore, our results provide a more robust and less uncertain representation of changes in the potential for compound flooding in Europe than previous projections.

Abstract Recent mass loss from ice sheets and ice shelves is now persistent and prolonged enough that it impacts downstream oceanographic conditions. To demonstrate this, we use an ensemble of coupled GISS‐E2.1‐G simulations forced with historical estimates of anomalous freshwater, in addition to other climate forcings, from 1990 through 2019. There are detectable differences in zonal‐mean sea surface temperatures (SST) and sea ice in the Southern Ocean, and in regional sea level around Antarctica and in the western North Atlantic. These impacts mostly improve the model's representation of historical changes, including reversing the forced trends in Antarctic sea ice. The changes in SST may have implications for estimates of the SST pattern effect on climate sensitivity and for cloud feedbacks. We conclude that the changes are sufficiently large that model groups should strive to include more accurate estimates of these drivers in all‐forcing historical simulations in future coupled model intercomparisons.

The global ocean is losing oxygen with warming. Observations and Earth system model projections, however, suggest that this global ocean deoxygenation does not equate to a simple and systematic expansion of tropical oxygen minimum zones (OMZs). Previous studies have focused on the Pacific Ocean; they showed that the outer OMZ deoxygenates and expands as oxygen supply by advective transport weakens, the OMZ core oxygenates and contracts due to a shift in the composition of the source waters supplied by slow mixing, and in between these two regimes oxygen is redistributed with little effect on OMZ volume. Here, we examine the OMZ response to warming in the Indian Ocean using an ensemble of Earth system model high-emissions scenario experiments from the Coupled Model Intercomparison Project Phase 6. We find a similar expansion–redistribution–contraction response but show that the unique ocean circulation pathways of the Indian Ocean lead to far more prominent OMZ contraction and redistribution regimes than in the Pacific Ocean. As a result, only the outermost volumes (oxygen>180 µmol kg−1) expand. The Indian Ocean experiences a broad oxygenation in the southwest driven by a reduction in waters supplied by the Indonesian Throughflow in favor of high-oxygen waters supplied from the southern Indian Ocean gyre. Models also project a strong localized deoxygenation in the northern Arabian Sea due to the rapid warming and shoaling of marginal sea outflows (Red Sea and Persian Gulf) and increases in local stratification with warming. We extend the existing conceptual framework used to explain the Pacific OMZ response to interpret the response in the Indian Ocean.

Coastal ecosystems release or absorb carbon dioxide (CO 2 ), methane (CH 4 ) and nitrous oxide (N 2 O), but the net effects of these ecosystems on the radiative balance remain unknown. We compiled a dataset of observations from 738 sites from studies published between 1975 and 2020 to quantify CO 2 , CH 4 and N 2 O fluxes in estuaries and coastal vegetation in ten global regions. We show that the CO 2 -equivalent (CO 2 e) uptake by coastal vegetation is decreased by 23–27% due to estuarine CO 2 e outgassing, resulting in a global median net sink of 391 or 444 TgCO 2 e yr −1 using the 20- or 100-year global warming potentials, respectively. Globally, total coastal CH 4 and N 2 O emissions decrease the coastal CO 2 sink by 9–20%. Southeast Asia, North America and Africa are critical regional hotspots of GHG sinks. Understanding these hotspots can guide our efforts to strengthen coastal CO 2 uptake while effectively reducing CH 4 and N 2 O emissions.

Modern climate projections lack adequate spatial and temporal resolution due to computational constraints. A consequence is inaccurate and imprecise predictions of critical processes such as storms. Hybrid methods that combine physics with machine learning (ML) have introduced a new generation of higher fidelity climate simulators that can sidestep Moore's Law by outsourcing compute-hungry, short, high-resolution simulations to ML emulators. However, this hybrid ML-physics simulation approach requires domain-specific treatment and has been inaccessible to ML experts because of lack of training data and relevant, easy-to-use workflows. We present ClimSim, the largest-ever dataset designed for hybrid ML-physics research. It comprises multi-scale climate simulations, developed by a consortium of climate scientists and ML researchers. It consists of 5.7 billion pairs of multivariate input and output vectors that isolate the influence of locally-nested, high-resolution, high-fidelity physics on a host climate simulator's macro-scale physical state. The dataset is global in coverage, spans multiple years at high sampling frequency, and is designed such that resulting emulators are compatible with downstream coupling into operational climate simulators. We implement a range of deterministic and stochastic regression baselines to highlight the ML challenges and their scoring. The data (https://huggingface.co/datasets/LEAP/ClimSim_high-res) and code (https://leap-stc.github.io/ClimSim) are released openly to support the development of hybrid ML-physics and high-fidelity climate simulations for the benefit of science and society.

GCM-Filters is a python package that allows scientists to perform spatial filtering analysis in an easy, flexible and efficient way.The package implements the filtering method based on the discrete Laplacian operator that was introduced by Grooms et al. (2021).The filtering algorithm is analogous to smoothing via diffusion; hence the name diffusion-based filters.GCM-Filters can be used with either gridded observational data or gridded data that is produced by General Circulation Models (GCMs) of ocean, weather, and climate.Spatial filtering of observational or GCM data is a common analysis method in the Earth Sciences, for example to study oceanic and atmospheric motions at different spatial scales or to develop subgrid-scale parameterizations for ocean models.

Abstract Global ocean oxygen loss is projected to persist in the future, but Earth system models (ESMs) have not yet provided a consistent picture of how it will influence the largest oxygen minimum zone (OMZ) in the tropical Pacific. We examine the change in the Pacific OMZ volume in an ensemble of ESMs from the CMIP6 archive, considering a broad range of oxygen (O2) thresholds relevant to biogeochemical cycles and ecosystems (5–160 µmol/kg). Despite OMZ biases in the historical period of the simulations, the ESM ensemble projections consistently fall into three regimes across ESMs: an expansion of low oxygenated waters (+0.8 [0.6, 1.0] × 1016 m3/century for O2 ≤ 120 µmol/kg, ESM median and interquartile range); a slight contraction of the OMZ core although more uncertain across ESMs (−0.1 [−0.5, 0.0] × 1016 m3/century for O2 ≤ 20 µmol/kg); and at the transition from contraction to expansion regimes, a spatial redistribution but near‐zero change in the volume of hypoxic waters (0.0 [−0.3, +0.1] × 1016 m3/century for O2 ≤ 60 µmol/kg). Changes in circulation and biology dictate the shift from expansion to contraction. Specifically, reduced subtropical ventilation controls the expansion of low oxygenated waters, while a combination of circulation and biological changes explains the contraction of the core (likely changes in mixing, reduced intermediate ventilation and oxygen demand). Increased model complexity (e.g., ecosystem dynamics and equatorial circulation) likely stabilize the OMZ response, suggesting that future changes might lie in the lower bound of current projections. The expansion of low oxygenated waters which delimit the optimum habitat of numerous marine species would severely impact ecosystems and ecosystem services.

Oceanic tracers are transported by oceanic motions of all scales, but only the large-scale motions are resolved by the present-day Earth system models. In these models, the unresolved lateral sub-gridscale tracer transport is generally parameterized through diffusive closures with a scale-independent diffusion coefficient. However, evidence from observations and theory suggests that diffusivity varies spatially and is length-scale dependent. Here we provide new scale-dependent quantification of the global surface diffusivities. To this end we use a recently developed statistical inversion method, MicroInverse, to diagnose horizontal surface diffusivities from observed sea surface temperature and idealized model simulation. We compare the results to theoretical estimates of mixing by the large-scale shear and by the sub-gridscale velocity fluctuations. The diagnosed diffusivity magnitude peaks in the tropics and western boundary currents with minima in the subtropical gyres (similar to 3000 and similar to 100 m(2) s(-1)) at similar to 40-km scale, respectively. Focusing on the 40-200-km length scale range, we find that the diffusivity magnitude scales with the length scale to a power n that is between 1.22 and 1.54 (90% confidence) in the tropics and also peaks at values above 1 in the boundary currents. In the midlatitudes we find that 0.58 < n < 0.87 (90% confidence). Comparison to the theory suggests that in regions with n > 1 the horizontal mixing is dominated by large-scale shear, whereas in regions where n < 1 the horizontal mixing is due to processes that are small compared to the 40-200-km length scale range considered in this study.

Global ocean oxygen loss - deoxygenation - is projected to persist in the future. Previous generations of Earth system models (ESMs) have, however, failed to provide a consistent picture of how deo...