Explore the vast range of titles available.

Subglacial drainage networks regulate the response of ice sheet flow to surface meltwater input to the subglacial environment. Simulating subglacial hydrology evolution is critical to projecting ice sheet sensitivity to climate, and contribution to sea‐level change. However, current numerical subglacial hydrology models are computationally expensive, and, consequently, evolving subglacial hydrology is neglected in large‐scale ice sheet simulations. We present a deep learning emulator of a state‐of‐the‐art subglacial hydrology model, trained at multiple Greenland glaciers. Our emulator performs strongly in both temporal (R2 > 0.99) and spatial (R2 > 0.95) generalization, offers high computational savings, and can be used to force numerical ice sheet models. This will enable century‐ and large‐scale ice sheet model simulations, including interactions between ice flow and increased meltwater input to the subglacial environment. Generally, our work demonstrates that machine learning can further improve ice sheet models, reduce computational bottlenecks, and exploit information from high‐fidelity models and novel observational platforms. Plain Language Summary Meltwater at the surface of ice sheets can drain to the subglacial environment, lubricate the bed, and influence ice sheet flow. Complex numerical subglacial hydrology models represent the subglacial drainage system, but are too computationally expensive to be included in large‐scale and long‐term ice sheet simulations. Consequently, model predictions of future ice sheet contribution to sea‐level rise ignore ice flow modulation by evolving subglacial hydrology. Here, we use deep learning to emulate a state‐of‐the‐art subglacial hydrology model. The emulator can directly force large‐scale ice sheet models to capture ice flow sensitivity to subglacial hydrology. The computational speed and accuracy of our emulator show the potential to use machine learning to efficiently incorporate previously neglected processes into ice sheet models. Key Points We develop a deep learning emulator to simulate evolving subglacial hydrology in response to meltwater input for ice sheet simulations The emulator shows generalization capabilities, large computational savings, and can be used to force numerical ice sheet models We demonstrate that machine learning has substantial potential in improving ice sheet models, through using information‐rich data sets

Antarctic supraglacial lakes (SGLs) have been linked to ice shelf collapse and the subsequent acceleration of inland ice flow, but observations of SGLs remain relatively scarce and their interannual variability is largely unknown. This makes it difficult to assess whether some ice shelves are close to thresholds of stability under climate warming. Here, we present the first observations of SGLs across the entire East Antarctic Ice Sheet over multiple melt seasons (2014–2020). Interannual variability in SGL volume is >200% on some ice shelves, but patterns are highly asynchronous. More extensive, deeper SGLs correlate with higher summer (December-January-February) air temperatures, but comparisons with modelled melt and runoff are complex. However, we find that modelled January melt and the ratio of November firn air content to summer melt are important predictors of SGL volume on some potentially vulnerable ice shelves, suggesting large increases in SGLs should be expected under future atmospheric warming. Antarctic supraglacial lakes (SGLs) have been linked to ice-shelf collapse and the subsequent acceleration of inland ice flow, but observations of SGLs remain relatively scarce and their interannual variability is largely unknown. This new study shows that lake area and volume vary substantially from year-to-year around the East Antarctic Ice Sheet and between ice shelves.

This study proposes a new lightning scheme applicable at the global scale, predicting lightning rates from climatic variables. Using satellite lightning records spanning a period of 29 years, we apply machine learning methods to derive a functional relationship between lightning and climate reanalysis data. In particular, we design a tree‐based regression scheme, representing different lightning regimes with separate single hidden layer neural networks of low dimensionality. We apply multiple complexity constraints in the development stages, which makes our lightning scheme straightforward to implement within global climate models (GCMs). We demonstrate that, for years not used for training, our lightning scheme captures 71.8%$71.8\\%$of the daily global spatio‐temporal lightning variability, which corresponds to a >43%${ >} 43\\%$relative improvement compared to well‐established lightning schemes. Similarly, the scheme correlates well with lightning observations for the monthly climatology (r>0.92)$(r > 0.92)$ , inter‐annual variability (r>0.76)$(r > 0.76)$ , and latitudinal and longitudinal distributions (r>0.87)$(r > 0.87)$ . Most notably, the lightning scheme brings a critical improvement in representing lightning magnitude and variability in the three tropical lightning chimney regions: central Africa, the Amazon, and the Maritime Continent. We implement the lightning scheme in the Community Earth System Model to verify its stability and performance as a GCM component, and we provide detailed implementation guidelines. As an intermediate approach between high‐dimensional machine learning models and first‐order lightning parameterizations, our lightning scheme offers GCMs a straightforward and efficient tool to improve lightning simulation, which is critical for representing atmospheric chemistry and naturally ignited wildfires. Plain Language Summary Lightning is a worldwide phenomenon, which affects atmospheric chemistry and can cause wildfire ignitions. However, representing lightning in climate models at the global scale is challenging, because it depends on small‐scale physical processes not explicitly represented in global models. In this study, we develop a lightning scheme, which estimates lightning rates from large‐scale climate variables. We use machine learning methods to extract differences in the relationship between lightning and climate in different lightning regimes. We impose several constraints to keep the lightning scheme simple, which facilitates implementation and use of the scheme as a component of global climate models. We show that predictions of the lightning scheme reproduce temporal and spatial variability of lightning observations. In addition, the match to observed lightning rates is improved compared to currently widely used lightning schemes. Key Points Using lightning records spanning 29 years, we develop a climate‐dependent lightning scheme applicable at the global scale Compared to established lightning schemes, spatio‐temporal variability is improved from daily to inter‐annual scales Our scheme employs machine learning with complexity constraints, enabling implementation in climate models and realistic lightning modeling

Models that simulate the evolution of polar firn are important for several applications in glaciology, including converting ice-sheet elevation change measurements to mass change and interpreting climate records in ice cores. We have developed the Community Firn Model (CFM), an open-source, modular model framework designed to simulate numerous physical processes in firn. The modules include firn densification, heat transport, meltwater percolation and refreezing, water isotope diffusion, and firn-air diffusion. The CFM is designed so that new modules can be added with ease. In this paper, we first describe the CFM and its modules. We then demonstrate the CFM's usefulness in two model applications that utilize two of its novel aspects. The CFM currently has the ability to run any of 13 previously published firn densification models, and in the first application we compare those models' results when they are forced with regional climate model outputs for Summit, Greenland. The results show that the models do not agree well (spread greater than 10 %) when predicting depth-integrated porosity, firn age, or the trend in surface elevation change. In the second application, we show that the CFM's coupled firn-air and firn densification models can simulate noble gas records from an ice core better than a firn-air model alone.

As surface melt is increasing on the Greenland Ice Sheet (GrIS), quantifying the retention capacity of the firn layer is critical to linking meltwater production to meltwater runoff. Firn-densification models have so far relied on empirical approaches to account for the percolation–refreezing process, and more physically based representations of liquid water flow might bring improvements to model performance. Here we implement three types of water percolation schemes into the Community Firn Model: the bucket approach, the Richards equation in a single domain and the Richards equation in a dual domain, which accounts for partitioning between matrix and fast preferential flow. We investigate their impact on firn densification at four locations on the GrIS and compare model results with observations. We find that for all of the flow schemes, significant discrepancies remain with respect to observed firn density, particularly the density variability in depth, and that inter-model differences are large (porosity of the upper 15 m firn varies by up to 47 %). The simple bucket scheme is as efficient in replicating observed density profiles as the single-domain Richards equation, and the most physically detailed dual-domain scheme does not necessarily reach best agreement with observed data. However, we find that the implementation of preferential flow simulates ice-layer formation more reliably and allows for deeper percolation. We also find that the firn model is more sensitive to the choice of densification scheme than to the choice of water percolation scheme. The disagreements with observations and the spread in model results demonstrate that progress towards an accurate description of water flow in firn is necessary. The numerous uncertainties about firn structure (e.g. grain size and shape, presence of ice layers) and about its hydraulic properties, as well as the one-dimensionality of firn models, render the implementation of physically based percolation schemes difficult. Additionally, the performance of firn models is still affected by the various effects affecting the densification process such as microstructural effects, wet snow metamorphism and temperature sensitivity when meltwater is present.

Decadal climate predictions are sensitive to model initialization and simulation of climate forced response and internal variability. Analogue-based initialization selects initial states matching observations from large climate model ensemble simulations, but neglects model performance. We implement performance-based model weighting, favoring models consistent with observations in climate forced response and stationary dynamics. Focusing on sea-surface temperature decadal predictions, we demonstrate the effectiveness of a deviance statistic, not previously used in model weighting schemes. We conduct performance-weighted predictions of pseudo-observations, targeting model realizations instead of observations, which show large decadal forecast potential skill improvement compared to unweighted predictions. We also find skill gains in decadal hindcasts of 95-year real-world sea-surface temperature observations, however at considerably lower levels. We explain this apparent contradiction by limited intrinsic predictability, similarity between unweighted and weighted ensembles, and inherent skill sampling uncertainties. Our analysis therefore highlights previously unrecognized challenges in validating performance-based model weighting for climate forecasting.

Perennial snow, or firn, covers 80 % of the Greenland ice sheet and has the capacity to retain surface meltwater, influencing the ice sheet mass balance and contribution to sea-level rise. Multilayer firn models are traditionally used to simulate firn processes and estimate meltwater retention. We present, intercompare and evaluate outputs from nine firn models at four sites that represent the ice sheet's dry snow, percolation, ice slab and firn aquifer areas. The models are forced by mass and energy fluxes derived from automatic weather stations and compared to firn density, temperature and meltwater percolation depth observations. Models agree relatively well at the dry-snow site while elsewhere their meltwater infiltration schemes lead to marked differences in simulated firn characteristics. Models accounting for deep meltwater percolation overestimate percolation depth and firn temperature at the percolation and ice slab sites but accurately simulate recharge of the firn aquifer. Models using Darcy's law and bucket schemes compare favorably to observed firn temperature and meltwater percolation depth at the percolation site, but only the Darcy models accurately simulate firn temperature and percolation at the ice slab site. Despite good performance at certain locations, no single model currently simulates meltwater infiltration adequately at all sites. The model spread in estimated meltwater retention and runoff increases with increasing meltwater input. The highest runoff was calculated at the KAN_U site in 2012, when average total runoff across models (±2σ) was 353±610 mm w.e. (water equivalent), about 27±48 % of the surface meltwater input. We identify potential causes for the model spread and the mismatch with observations and provide recommendations for future model development and firn investigation.

Variability in oceanic conditions directly impacts ice loss from marine outlet glaciers in Greenland, influencing the ice sheet mass balance. Oceanic conditions are available from Atmosphere‐Ocean Global Climate Model (AOGCM) output, but these models require extensive computational resources and lack the fine resolution needed to simulate ocean dynamics on the Greenland continental shelf and close to glacier marine termini. Here, we develop a statistical approach to generate ocean forcing for ice sheet model simulations, which incorporates natural spatiotemporal variability and anthropogenic changes. Starting from raw AOGCM ocean heat content, we apply: (a) a bias‐correction using ocean reanalysis, (b) an extrapolation accounting for on‐shelf ocean dynamics, and (c) stochastic time series models to generate realizations of natural variability. The bias‐correction reduces model errors by ∼25% when compared to independent in‐situ measurements. The bias‐corrected time series are subsequently extrapolated to fjord mouth locations using relations constrained from available high‐resolution regional ocean model results. The stochastic time series models reproduce the spatial correlation, characteristic timescales, and the amplitude of natural variability of bias‐corrected AOGCMs, but at negligible computational expense. We demonstrate the efficiency of this method by generating >6,000 time series of ocean forcing for >200 Greenland marine‐terminating glacier locations until 2100. As our method is computationally efficient and adaptable to any ocean model output and reanalysis product, it provides flexibility in exploring sensitivity to ocean conditions in Greenland ice sheet model simulations. We provide the output and workflow in an open‐source repository, and discuss advantages and future developments for our method. Plain Language Summary Model simulations of the Greenland ice sheet (GrIS) require knowledge of ocean conditions. The evolution of ocean conditions has a strong impact on ice sheet model predictions, as there are more than 200 glaciers in Greenland flowing directly into the ocean. However, modeling oceanic forcing is difficult. The state‐of‐the‐art approach is to use output from Atmosphere‐Ocean Global Climate Models (AOGCMs). But these models cannot accurately capture the ocean dynamics on the Greenland shelf, and they can show strong biases compared to observations. Furthermore, AOGCMs are computationally expensive, meaning that it is impossible to thoroughly characterize the uncertainty associated with the chaotic nature of climate. Here, we propose a procedure to bias‐correct and extrapolate oceanic output from AOGCMs. Our method exploits observational datasets, as well as available high‐resolution ocean model results. Using statistical models, we reproduce patterns of spatiotemporal ocean variability at low computational expense, and represent internal climate variability and global warming trends. The goal is to provide a scalable procedure to generate ocean forcing for long‐term GrIS model predictions. Key Points We develop a statistical method to generate ocean forcing boundary conditions for Greenland ice sheet model simulations The method bias‐corrects and extrapolates global climate model output using reanalysis products and high‐resolution model results Stochastic time series models reproduce the spatiotemporal variability of ocean conditions at negligible computational expense

Regional climate variability manifests through distinct atmospheric regimes influencing weather, climate, and predictability. Yet, their response to global warming remains unresolved. Using a hundred realisations from the Community Earth System Model Large Ensemble, we examine shifts in wintertime North Atlantic atmospheric regimes and the North Atlantic Oscillation before and after 1995, highlighting the detectable influence of anthropogenic warming on atmospheric circulation. The large-ensemble framework isolates internal variability by removing the ensemble mean. Under anthropogenic warming, the number of regime states associated with the forced response remains constant, although their spatial circulation patterns undergo substantial reorganisation. In contrast, internal variability alone exhibits a reduction in the number of regime states. Future projections indicate a shift toward more frequent positive phases of the North Atlantic Oscillation, accompanied by low-amplitude negative phases late in the century, alongside a marked decline in its variability and altered mid-tropospheric westerlies. Anthropogenic warming reorganises North Atlantic atmospheric circulation patterns, producing a shift towards more positive North Atlantic Oscillation phases, according to analysis of climate model simulations.

Most climatic and glaciological processes exhibit internal variability, which is omitted from many ice sheet model simulations. Prior studies have found that climatic variability can change ice sheet sensitivity to the long-term mean and trend in climate forcing. In this study, we use an ensemble of simulations with a stochastic large-scale ice sheet model to demonstrate that variability in frontal ablation of marine-terminating glaciers changes the mean state of the Greenland Ice Sheet through noise-induced drift. Conversely, stochastic variability in surface mass balance does not appear to cause noise-induced drift in these ensembles. We describe three potential causes for noise-induced drift identified in prior statistical physics literature: noise-induced bifurcations, multiplicative noise, and nonlinearities in noisy processes. Idealized simulations and Reynolds decomposition theory show that for marine ice sheets in particular, noise-induced bifurcations and nonlinearities in variable ice sheet processes are likely the cause of the noise-induced drift. We argue that the omnipresence of variability in climate and ice sheet systems means that the state of real-world ice sheets includes this tendency to drift. Thus, the lack of representation of such noise-induced drift in spin-up and transient ice sheet simulations is a potentially ubiquitous source of bias in ice sheet models.