The first stage fits the mean, variance, skewness and kurtosis of temperature as smooth functions of day of year. It combines Generalized Additive Models for Location, Scale and Shape with a Sinh-Arcsinh, or SHASH, distribution, whose tail-weight parameter provides a way to represent the observed platykurtosis. The resulting quantile residuals were close to the ideal standard-normal benchmarks of mean 0, standard deviation 1, skewness 0 and kurtosis 3. GSAT’s Akaike information criterion was 78.4 units better than that of Gaussian GAMLSS, a difference the authors describe as decisive by information-criterion standards: the corresponding relative likelihood for the Gaussian specification was approximately 10^−17. This stage also produced the largest gains in the Monte Carlo comparison, reducing standard-deviation root-mean-square error by 80.5% and skewness RMSE by 26.7%, while kurtosis RMSE was 11.6% higher.

The second stage accounts for dependence among days through an autoregressive moving-average model of order (3,1), while the third uses an exponential generalized autoregressive conditional heteroscedasticity model of order (1,1) to represent changing residual volatility. The authors reported a possible warm-anomaly leverage asymmetry, but presented it with bootstrap uncertainty and did not treat it as a main source of performance. Across 1,000 Monte Carlo replications, GSAT beat the Fourier–ARMA–Normal benchmark on six of eight measures. Component ablation experiments attributed the improvement primarily to the stages that modeled seasonal variance and distributional shape, and leave-one-year-out testing supported out-of-sample generalization.

Evidence from six UAE stations—Abu Dhabi, Dubai, Sharjah, Al Ain, Ras Al Khaimah and Fujairah—broadened the assessment beyond the five-year Abu Dhabi record. The station records covered 13 to 39 complete years, five of the country’s seven emirates and both coastlines. Every station showed a highly significant seasonal variance cycle, with Bartlett p-values below 10^−55 and below 10^−200 at the four longest-record stations. The four-moment GAMLSS-SHASH specification improved on Gaussian GAMLSS by 73 to 462 AIC units at the stations. Tail weight also varied by season and location, indicating that UAE temperatures are not uniformly platykurtic. A peaks-over-threshold analysis was consistent with a bounded, or light, upper tail, although both GSAT and the benchmark overestimated the maximum length of warm spells.

Further validation used temporally blocked cross-validation across 31 test blocks, each containing as many as five complete years, with eight to 35 training years per fold. Four folds required strict forward testing. Held-out monthly-climatology Nash–Sutcliffe efficiencies ranged from 0.947 to 0.995, while ensemble-mean daily-scale values ranged from 0.85 to 0.93; monthly-mean mean absolute percentage error was 1.2–3.5%. GSAT had lower held-out standard-deviation RMSE than the identically fitted benchmark in 30 of 31 folds. Its synthesis stage also mapped the simulated latent process through the fitted residuals’ empirical cumulative distribution function, with probability-integral-transform calibration verified directly. Synthetic series reproduced the observed autocorrelation of anomalies and squared anomalies. The study presents GSAT as a way to generate synthetic temperature series while preserving not only seasonal averages but also evolving variability, skewness, kurtosis and short-range temporal structure; its validation is regional and observational rather than evidence of operational forecasting skill.

Subject of Research: Evaluation of GSAT for generating synthetic daily temperature series in the United Arab Emirates

Article Title: Gsat: a higher-order stochastic framework for synthetic daily temperature generation in the United Arab Emirates

Article References: Elseidi, M. (2026). Gsat: a higher-order stochastic framework for synthetic daily temperature generation in the United Arab Emirates. Theoretical and Applied Climatology, 157(11), Article 738. https://doi.org/10.1007/s00704-026-06666-1

Image Credits: AI Generated

DOI: 10.1007/s00704-026-06666-1

Keywords: synthetic weather generation, daily temperature, United Arab Emirates, GAMLSS-SHASH, stochastic modeling, seasonal variance, ARMA models, eGARCH, cross-validation, climate extremes, Gsat, higher-order

The research, published in Nature Communications by Catherine H. Gregory, Urs Hofmann Elizondo, Thibault Guinaldo, and Thomas L. Frölicher, uses a technique known as ensemble boosting to explore physically plausible alternative versions of the 2023 event. Rather than asking what did happen, the team asked what could have happened under the same large-scale climate conditions. The answer is sobering. In their boosted simulations, marine heatwaves emerged that lasted up to 579 days, reached maximum intensities of 4.5 degrees Celsius above the long-term climatology, and accumulated a total heat stress of 717 degree-days. By comparison, the observed 2023 event lasted roughly 240 days, peaked at about 2.9 degrees Celsius above normal, and accumulated 224 degree-days of cumulative intensity.

Ensemble boosting is a relatively new approach in the study of extreme events, and it addresses a persistent problem in climate science: state-of-the-art Earth system models are run as large collections of simulations, but only a handful of members will ever reproduce a truly rare extreme. The technique works by taking a coupled Earth system model and selectively perturbing the initial conditions of individual model members during periods that resemble the preconditioning of a real event. These perturbations are tiny, physically consistent nudges that represent the natural chaos of the atmosphere and ocean. If a nudged simulation goes on to produce an extreme event, that event is not an artifact; it is a legitimate alternative evolution of the same climate system, one that the dice of weather variability could plausibly have rolled.

Applied to the North Atlantic in 2023, the method produced a family of storylines that bracket the observed event and extend well beyond it. The most extreme simulated heatwaves were not freak occurrences requiring exotic physics. They arose from the same oceanic and atmospheric ingredients that produced the real event, arranged in somewhat less favorable configurations. This is the central message of the study: the margin between the 2023 heatwave we experienced and a substantially more severe outcome was narrower than most people would assume, and it was governed largely by the luck of day-to-day and season-to-season weather.

The detailed analysis of the longest and most intense boosted events reveals a set of contrasting mechanisms that conspire to build and sustain extreme ocean warmth. A central player is persistent high atmospheric pressure over the Euro-Atlantic region. When a high-pressure system settles over the North Atlantic, it suppresses cloud formation and calms the winds. Calmer winds mean less evaporative cooling, so the ocean loses less heat to the atmosphere through turbulent fluxes. Clearer skies, meanwhile, allow more incoming shortwave radiation to penetrate the sea surface and warm the upper ocean directly. The combination of reduced turbulent heat loss and enhanced solar heating acts like a lid and a magnifying glass at once, allowing heat to accumulate in the surface layers week after week.

The third mechanism identified by the researchers involves the redistribution of heat within the upper ocean. Marine heatwaves are not simply a story of local heating; they also depend on how heat is stirred, mixed, and transported. In the most extreme simulated events, changes in ocean circulation and mixing patterns kept warm water trapped near the surface or advected additional warmth into the affected region, reinforcing the surface anomaly. The interplay between these atmospheric drivers and oceanic redistribution helps explain why some boosted events persisted for well over a year and a half, far outlasting the observed event, while others of comparable peak intensity faded more quickly.

These findings matter because the consequences of a marine heatwave scale with both its intensity and its duration. Peak temperature anomalies determine the immediate physiological stress on organisms such as corals, fish larvae, and plankton, while cumulative intensity, often measured in degree-days, captures the prolonged exposure that can push ecosystems past tipping points. A heatwave lasting 579 days with a cumulative intensity more than three times that of the observed 2023 event would represent a fundamentally different level of stress for North Atlantic ecosystems, fisheries, and the communities that depend on them. The boosted storylines therefore function as a stress test, revealing the upper range of what present-day climate variability can deliver.

The study also places these present-day possibilities in the context of future warming, and this is where the implications become most striking. The researchers show that events comparable to the observed 2023 heatwave, which today still qualifies as an extreme outlier, become increasingly common as global warming progresses. In a warmer climate, the baseline temperature of the North Atlantic rises, so the same weather patterns that produced the 2023 anomaly would operate on a warmer starting point, pushing temperatures into ranges that would today seem almost implausible. In effect, the boosted ensemble offers a preview of the kinds of events that routine greenhouse warming will make ordinary. What was a near-miss in 2023 could, within decades, fall within the range of typical summer conditions.

The North Atlantic holds particular importance in the Earth system, influencing weather over Europe and North America, absorbing substantial amounts of heat and carbon dioxide, and hosting some of the most productive fisheries in the world. Prolonged extreme warmth in this basin can alter stratification and nutrient supply, shift the distribution of commercially important species, and contribute to atmospheric circulation anomalies that affect terrestrial heatwaves and droughts. The 2023 event already prompted intense scientific debate about its causes, with proposed explanations ranging from reduced aerosol pollution from shipping to El Niño development and unusual atmospheric circulation. This new study does not settle that debate, but it reframes it: whatever the underlying preconditioning was, the weather-dependent upper tail of possible outcomes was much more severe than what materialized.

Methodologically, the work demonstrates the growing value of storyline-based approaches to climate risk. Instead of assigning a single probability to an observed extreme, ensemble boosting generates a spectrum of physically self-consistent alternatives that illuminate which processes control an event’s severity and how close reality came to a worse outcome. For decision-makers in fisheries management, coastal planning, and marine conservation, such storylines are arguably more useful than abstract return periods, because they describe concrete scenarios that could plausibly unfold next time the same preconditioning arises. As computing resources expand and coupled models improve, techniques of this kind are likely to become a standard tool for assessing the risks posed by compound and record-shattering extremes, on land as well as at sea.

The broader takeaway is a reminder that the record books of the last few years should not be read as a ceiling. The 2023 North Atlantic marine heatwave was extraordinary, but the physics of the present-day climate system, probed systematically by the boosted simulations of Gregory and colleagues, permits events that are longer, hotter, and more cumulative in their impact. As global warming continues to raise the baseline, the gap between what is observed and what is possible will narrow from the other direction, with once-hypothetical extremes becoming the new benchmark. Preparing for the worst plausible outcome, rather than merely the worst recorded one, is emerging as an essential principle for understanding and adapting to the changing ocean.

Subject of Research: Alternative climate model scenarios of the 2023 North Atlantic marine heatwave and its potential for greater extremity

Article Title: The 2023 North Atlantic marine heatwave could have evolved into a far more extreme event

Article References: Gregory, C. H., Elizondo, U. H., Guinaldo, T., & Frölicher, T. L. (2026). The 2023 North Atlantic marine heatwave could have evolved into a far more extreme event. Nature Communications. https://doi.org/10.1038/s41467-026-78863-5

Image Credits: AI Generated

DOI: 10.1038/s41467-026-78863-5

Keywords: marine heatwave, North Atlantic, ensemble boosting, Earth system model, ocean warming, climate extremes, sea surface temperature, Euro-Atlantic high pressure, cumulative intensity, climate change, physical oceanography, future warming

The research team applied a machine learning framework built around Shapley Additive Explanations, a technique borrowed from game theory that assigns each input variable a quantifiable share of a model’s output. Paired with species distribution models, this approach allowed the scientists to decompose projected changes in mammal habitat suitability into the individual contributions of 16 distinct environmental drivers. These drivers span climate averages, such as mean annual temperature and precipitation; climate variability, including seasonal fluctuations; climate extremes, such as heatwaves and droughts; and land cover characteristics, from forest extent to human-modified landscapes. Rather than treating the environment as a single blended force, the method reveals the fingerprint of each driver separately, across space and through time.

What emerges from this decomposition is a picture of ecological change far more nuanced than conventional global assessments have offered. The team introduced two novel diagnostic measures to interpret their results. The first, which they call directional turnover, identifies locations where a given driver flips from being favorable to unfavorable for a species, or the reverse. The second, favorability shift, captures places where a driver’s influence strengthens or weakens without changing its sign. Together, these diagnostics expose the geography of ecological risk and opportunity in a way that aggregate vulnerability metrics cannot.

One of the study’s most striking findings concerns climate extremes. Heatwaves, cold snaps, and droughts impose intense pressures on mammal populations, but those pressures are spatially restricted, concentrated in specific regions rather than spread uniformly across continents. This means that while extreme events may not affect the largest total area, the severity of their local impact can be profound, pushing species in those hotspots toward the edges of their physiological tolerance. Meanwhile, temperature-related drivers dominate the overall pattern of exposure, meaning that gradual warming trends shape the broad geography of risk more than any other factor the team measured.

Perhaps the most consequential insight, however, is that the same environmental driver can act as a threat in one region while creating opportunity in another. Rising temperatures that push a mammal beyond its thermal limits in the tropics may simultaneously open previously inhospitable habitat at higher latitudes. Precipitation changes that desiccate one ecosystem may green another. This spatial divergence has been largely invisible in traditional assessments that average effects across broad regions, and it carries direct implications for conservation planning. A species flagged as vulnerable based on range-wide averages may actually be gaining ground in parts of its distribution while collapsing in others, and interventions designed around aggregate metrics may misallocate scarce resources.

The framework also revealed that specialist species, those with narrow ecological tolerances, are disproportionately exposed to temperature-related drivers compared with generalists. This aligns with long-standing ecological theory, which predicts that specialists should be more sensitive to environmental change because they lack the behavioral and physiological flexibility to buffer against it. The fact that the explainable AI approach independently recovered this expected pattern serves as a check on the method’s plausibility, even though the authors are careful to note that their framework assesses whether projected spatial patterns align with ecological expectations rather than establishing strict causation.

Technically, the strength of the Shapley-based approach lies in its consistency and local accuracy. For any single prediction, the Shapley values of all input variables sum exactly to the difference between the model’s output and a baseline expectation, ensuring that contributions are neither double-counted nor lost. By computing these values at every grid cell across the globe and for every species, the team generated maps of driver-specific influence that can be compared across taxa, regions, and time periods. This turns what was previously a global summary statistic into a spatially explicit diagnostic tool, one that conservation practitioners can interrogate at the resolution relevant to on-the-ground decisions.

The implications extend beyond mammals. The analytical framework is fundamentally agnostic about taxonomic group; it could be applied to birds, amphibians, plants, or freshwater organisms, provided suitable occurrence data and environmental layers are available. As climate datasets grow finer in resolution and land cover products become more accurate, the same machinery could resolve driver contributions at scales relevant to individual protected areas or wildlife corridors. The authors suggest that their approach advances biodiversity forecasting toward a more actionable understanding of ecological risk and opportunity, shifting the conversation from how much change is coming to precisely where, why, and in which direction it will unfold.

For the 1,992 mammal species included in the analysis, the message is sobering but constructive. Climate change is not a uniform tide lifting or sinking all habitats equally. It is a mosaic of localized pressures and unexpected openings, shaped by the interplay of averages, variability, extremes, and land transformation. By making the invisible architecture of these forces visible, explainable machine learning offers conservation science something it has long needed: a way to see not just that biodiversity is changing, but exactly which lever of global change is pulling hardest, and where pulling back could make the greatest difference.

Subject of Research: Explainable machine learning analysis of how climate and land cover drivers reshape future distributions of terrestrial mammals

Article Title: Explainable artificial intelligence reveals spatially divergent effects of global change on mammals

Article References: Song, L., Frazier, A. E., Kedron, P., Araujo, D. S. A., Cui, D., Enquist, B. J., Hinojo-Hinojo, C., Krieger, J. M., Maitner, B. S., Merow, C., Moulatlet, G. M., Nikolopoulos, E. I., & Roehrdanz, P. R. (2026). Explainable artificial intelligence reveals spatially divergent effects of global change on mammals. Nature Communications. https://doi.org/10.1038/s41467-026-78677-5

Image Credits: AI Generated

DOI: 10.1038/s41467-026-78677-5

Keywords: explainable AI, Shapley values, species distribution models, climate change, mammals, biodiversity, climate extremes, land cover, ecological forecasting, conservation, biogeography, Nature Communications

The research team, led by Xiaoyang Li of the University of Tokyo’s Institute of Industrial Science together with Kei Yoshimura and Hironori Fudeyasu of Yokohama National University, built a national-scale impact-based flood-damage assessment framework around the Integrated Land Simulator, or ILS. This platform couples two established models: MATSIRO, a physical land surface model that handles soil, canopy, and runoff processes, and CaMa-Flood, a hydrodynamic river routing model that simulates floodplain inundation using the Stabilized Local Inertial Equation, an adaptive time step scheme, and an explicit levee and reservoir scheme. The modeling chain was anchored in high-resolution topography from the MERIT digital elevation model, allowing inundation depths computed at roughly 1.5-kilometer resolution to be downscaled to about 30 meters, fine enough to resolve individual buildings and neighborhoods.

The atmospheric side of the experiment relied on the Weather Research and Forecasting Model, version 4.3.3, run in a one-way nested configuration with an outer domain at 15-kilometer resolution and an inner domain at 5 kilometers, both covering the Japanese archipelago and surrounding seas with 48 vertical layers. Initial and boundary conditions came from the Japanese 55-year Reanalysis, and the physics suite included the RRTMG radiation schemes, the WSM6 cloud microphysics parameterization, the Kain-Fritsch cumulus scheme, and the Yonsei University planetary boundary layer scheme. The crucial twist was the perturbation method: the researchers systematically shifted the reanalysis fields longitudinally in increments of 0.2 degrees, spanning 11 degrees westward to 5 degrees eastward. In total, 81 plausible versions of Typhoon Hagibis were simulated, each following a slightly different east-west path across Japan.

Before running these scenarios, the team validated the control simulation against observations. The WRF control run reproduced the spatial distribution and intensity of accumulated rainfall recorded by the Japan Meteorological Agency’s radar network during the storm’s passage, with heavy rainfall concentrated in central Japan and along the Pacific coast, exactly as observed. Downstream, the ILS control simulation matched observed maximum river discharge and river water levels from the ministry’s gauging network remarkably well, with correlation coefficients of 0.90 for discharge and 0.80 for water level. The validation also revealed something important: even the small westward and eastward biases in the simulated storm track produced corresponding shifts in simulated rainfall and river response, a first hint of how sensitive the entire flood system is to the precise position of the typhoon.

The headline finding is stark. A longitudinal shift of just 0.2 degrees, roughly 20 kilometers at Japan’s latitude, produced an average difference in total flood damage of 83.3 billion yen between adjacent scenarios. The response was highly nonlinear. In one dramatic case, moving the storm from the w012 to the w014 scenario, a shift of only 0.2 degrees westward, increased national damage by 956.8 billion yen, because the altered track steered intense rainfall and severe inundation directly onto the densely populated Osaka and Nagoya metropolitan areas. Across the full ensemble, total damage generally rose as the storm moved eastward from the far-west scenarios, with two pronounced peaks: 1,966.3 billion yen in the e010 simulation and 1,522.6 billion yen in the w014 simulation, after which damage fell off rapidly.

The geography of worst-case tracks differs sharply by region, and that is where the study becomes a planning tool rather than a curiosity. The single most damaging scenario for Japan as a whole, e010, placed the typhoon’s landfall over the Kanto Plain rather than the mountainous interior, unleashing intense rainfall and widespread inundation across greater Tokyo, the country’s most populous metropolitan area. In that scenario, Kanto alone suffered 1,571.7 billion yen in damage. For the Kinki region around Osaka, the worst track was w016, producing 911.4 billion yen, while for Chubu around Nagoya it was w004, producing 642.2 billion yen. In the control run, total damage reached about 1,181.1 billion yen, with 724.1 billion yen concentrated in Kanto. A modest 0.2-degree westward shift cut the national total to 897.9 billion yen but multiplied Kinki’s damage more than eightyfold, from 9.3 billion to 752.5 billion yen, as the rainfall belt migrated toward Osaka, Chugoku, and Shikoku.

Most of the simulated damage, like most of the real damage, occurred in floodplain areas along the Pacific coast of Kanto, Chubu, Kinki, Shikoku, and Tohoku, with Kyushu’s damage concentrated instead along the Japan Sea coast. This pattern reflects the fact that 96.8 percent of Hagibis’s actual flood damage came from inland river overflows rather than coastal storm surge, which justified the study’s focus on riverine flooding. Because these floodplains coincide with Japan’s highest population densities, the authors argue that enhanced flood protection infrastructure and impact-based early warning systems are urgently needed there. The framework also quantified the value of hard infrastructure: in levee-height sensitivity experiments ranging from a 20 percent decrease to a 20 percent increase, total damage fell from 1,189.1 billion yen to 1,152.9 billion yen, an average reduction of about 9.1 billion yen for every 10 percent increase in levee height.

Technically, the damage estimates were built from the ground up. Building areas came from the global three-dimensional building footprint dataset, and asset values per unit floor area were drawn from Japan’s 2018 Housing and Land Survey, set at 37.05 thousand yen per square meter for structural assets and 91.78 thousand yen per square meter for contents. Downscaled inundation depths were converted into damage rates using the empirical depth-damage relationships in the ministry’s Manual for Economic Evaluation of Flood Control Investment. The authors are careful to note the limits: land values and indirect losses such as supply-chain disruptions, business interruptions, and long-term recovery costs were excluded, so the figures represent a lower bound on the true economic impact. The study is also explicitly retrospective and scenario-based, not an operational real-time forecasting system for evacuation decisions.

Even so, the implications reach well beyond one storm. The researchers point out that forecast errors of a similar magnitude actually occurred with Hagibis, with a notable westward bias in track forecasts roughly three days before landfall, meaning the kind of uncertainty explored here is not hypothetical. As the climate warms and the hydrological cycle accelerates, extreme rainfall events across East Asia are expected to intensify, while aging populations in Japan, China, and South Korea make evacuation harder; 77 percent of Hagibis’s victims, 67 of 87 people, were aged 60 or older. By translating small meteorological uncertainties into concrete yen-denominated damage figures for specific regions and cities, the framework gives decision-makers a way to weigh investments in levees, warning systems, and preparedness against the plausible worst cases. It also fills a genuine gap: national-scale impact-based flood forecasting has been demonstrated in East Africa and the Philippines, but never before in East Asia, and the authors suggest this integrated approach of atmospheric modeling, hydrological simulation, and socioeconomic exposure data could become a template for flood-risk management across the region.

Subject of Research: Sensitivity of national-scale flood damage in Japan to typhoon-track shifts and levee height during Typhoon Hagibis (2019)

Article Title: Impact-based flood-damage assessment in Japan under longitudinal typhoon-track perturbations and levee-height variations during Hagibis (2019)

Article References: Li, X., Yoshimura, K., & Fudeyasu, H. (2026). Impact-based flood-damage assessment in Japan under longitudinal typhoon-track perturbations and levee-height variations during Hagibis (2019). Communications Earth & Environment, 7(1), Article 787. https://doi.org/10.1038/s43247-026-04077-8

Image Credits: AI Generated

DOI: 10.1038/s43247-026-04077-8

Keywords: Typhoon Hagibis, flood damage, Japan, typhoon track, levees, impact-based forecasting, hydrology, inundation modeling, WRF, disaster risk, Kanto, climate extremes

The research, led by Aytaç Paçal of the German Aerospace Center (DLR) together with colleagues at the University of Bremen, Universidad Carlos III de Madrid, the Universitat de València, and DLR, centers on a Variational Autoencoder, or VAE. This type of neural network, first developed as a generative model, works by compressing enormous, high-dimensional inputs into a much smaller latent space and then reconstructing the original data from that compressed code. In this case, the inputs were not photographs but spatiotemporal snapshots of the atmosphere: nine climate variables, including temperature, humidity, wind components, mean sea level pressure, geopotential height at 500 hPa, cloud cover, stream function, and surface solar radiation, tracked over eleven-day windows across the North Atlantic and Europe.

To build those inputs, the team started with the ERA5 reanalysis dataset, a continuous reconstruction of global weather stretching back to 1940. Heatwave onsets were identified over Western European land areas by comparing daily maximum 2-meter temperatures against local 90th percentile thresholds, calculated for each calendar day using a fifteen-day moving window from the 1941 to 1980 baseline. A density-based clustering algorithm, GDBSCAN, then stitched spatially and temporally contiguous hot grid cells into discrete events, yielding 2,565 unique heatwave onset dates between 1941 and 2022. Each onset was wrapped in a window spanning five days before and five days after, so the model could learn not just what a heatwave looks like at its peak, but how the atmosphere builds toward it and breaks down afterward.

The VAE was deliberately trained on the historical period from 1941 to 1990, then evaluated on recent events from 2001 to 2022. This historical-to-recent split creates what the authors call an out-of-distribution setting: a test in which the model must represent climate conditions that differ from anything it learned. The architecture relied on three-dimensional convolutions, which process space and time simultaneously, proving more stable than hybrid designs combining two-dimensional convolutions with LSTM layers. The model achieved reconstruction skill, measured by R-squared scores of roughly 0.72 to 0.76 across training, validation, and test sets, with the strongest performance for the large-scale circulation fields such as geopotential height and sea level pressure, and lower skill for the noisier near-surface winds.

When the researchers encoded recent heatwaves into the latent space and visualized the results, a striking pattern emerged. Events from 2001 to 2022 occupied a ring-like structure with fewer samples at the center and a pronounced accumulation toward one side of the space, regions largely untouched by the historical training events. Statistical tests quantified the drift: using Mahalanobis distance to measure how far the recent distribution had moved, summer samples shifted by D equal to 2.00, with about 83 percent of that signal attributable to thermodynamic warming. When the linear warming trend was removed from the data, statistically significant shifts persisted across all four seasons, and winter emerged as the season with the largest remaining dynamical displacement, at D equal to 1.40. In other words, beyond simple mean warming, the multivariate atmospheric structure of heatwaves itself appears to be evolving.

Clustering the latent space with a Gaussian Mixture Model revealed that four components best describe the test-period heatwaves, and each cluster mapped onto a distinct physical regime. One cluster captured winter warm spells, driven by persistent anticyclonic blocking centered over the British Isles and Western Europe, with suppressed cloud cover, warm advection from the Atlantic, and a structure resembling the omega-block configurations known from the literature. A second, dominant summer cluster showed the classic signature of intense European heat: strong positive temperature, humidity, and geopotential height anomalies, a deep North Atlantic trough feeding a ridge over Europe, and near-stagnant winds under cloud-free skies. Two further clusters both captured transition-season events in spring and autumn, yet displayed nearly opposite anomaly patterns, one ruled by Greenland blocking and enhanced solar radiation, the other by stronger pressure anomalies and warm-air advection.

A crucial validation step came from testing the model against some of the most infamous heatwaves on record. The team located the 2003 Western European event, the 2010 Russian heatwave, and the 2018 Scandinavian heatwave within the learned latent space and constructed composite anomaly maps from each event’s ten nearest neighbors. The resulting 500 hPa geopotential height anomalies reproduced the well-documented circulation signatures of these benchmark events: a ridge over central Europe in 2003, strong positive height anomalies over western Russia in 2010, and Scandinavian height anomalies in 2018. The fact that an unsupervised model, trained without any labels or human guidance, independently organized these events in a physically consistent way underscores the promise of the method.

The findings carry weight because heatwaves are intensifying faster in Western Europe than climate models have generally projected, with recent studies showing that unprecedented extreme heat has grown roughly tenfold and that economic losses from events like those of 2003, 2010, 2015, and 2018 reached 0.3 to 0.5 percent of European gross domestic product. By capturing the interplay between circulation, humidity, radiation, and temperature, the VAE approach offers a window into why: enhanced land-atmosphere feedbacks, moisture limitations, and shifting circulation regimes can amplify and restructure heat extremes beyond what a thermodynamic trend alone would predict. The authors caution, however, that linear detrending cannot fully separate forced changes from natural variability, and that the identified clusters depend on the chosen variables, thresholds, and model hyperparameters.

The team also probed an open question with direct forecasting implications: do the pre-onset circulation patterns identified by the model appear during quiet periods that never produce a heatwave? A preliminary comparison against ordinary summer days from 2015 to 2021 found no clear predictive signal, suggesting these large-scale configurations can occur without always triggering extreme heat. Future work, the researchers suggest, could add soil moisture, vegetation indices, or apply the same framework to CMIP6 climate simulations to see how well models capture historical heatwave patterns and where their biases lie.

What makes this study resonate far beyond the technical community is its central message: heatwaves are not simply hotter versions of the past, but events whose underlying atmospheric machinery is shifting. By letting a machine discover those shifts without labels, thresholds, or preconceived categories, the researchers have produced a data-driven atlas of heatwave regimes that aligns with established meteorology yet reveals changes invisible to traditional single-variable methods. As Europe braces for hotter summers, warmer winters, and increasingly erratic shoulder-season extremes, tools like this VAE may become essential for detecting when the atmosphere’s playbook has changed, and for building the climate risk assessments on which adaptation strategies depend.

Subject of Research: Multivariate atmospheric drivers of Western European heatwaves analyzed with unsupervised machine learning

Article Title: A multivariate analysis of atmospheric drivers for Western European heatwaves

Article References: Paçal, A., Hassler, B., Weigel, K., Fernández-Torres, M.-Á., Camps-Valls, G., & Eyring, V. (2026). A multivariate analysis of atmospheric drivers for Western European heatwaves. Earth System Dynamics, 17(4), 955-986. https://doi.org/10.5194/esd-17-955-2026

Image Credits: AI Generated

DOI: 10.5194/esd-17-955-2026

Keywords: heatwaves, Western Europe, Variational Autoencoder, machine learning, ERA5, atmospheric blocking, climate extremes, geopotential height, latent space, Earth System Dynamics, unsupervised learning, climate change

The research team evaluated an ensemble of nine fully coupled simulations from the Mediterranean Coordinated Regional Downscaling Experiment, known as Med-CORDEX. These Regional Climate System Models, or RCSMs, are nested within parent Global Climate Models from the CMIP5 and CMIP6 archives but operate at dramatically finer scales: ocean grids of roughly 6 to 30 kilometers and atmospheric grids of 12 to 50 kilometers, compared with the coarse one-degree ocean resolution typical of global models. That finer grid matters because the Mediterranean is a mosaic of narrow straits, semi-enclosed sub-basins, and intense regional wind systems such as the Mistral and Bora, none of which a global model can properly resolve. The team compared daily sea surface temperature fields and derived marine heatwave properties against satellite observations spanning 1982 to 2020, asking a deceptively simple question: does the added computational complexity of regional downscaling actually translate into skill?

The answer, published on 9 September 2026, is nuanced and depends heavily on which metric you examine. The researchers quantified added value using an index that measures whether a regional model reproduces the observed probability distribution of a variable more accurately than its parent global model, with statistical significance tested through bootstrap randomization at every grid point. For sea surface temperature, the results were split. Seven of the nine regional models improved the spatial pattern of mean sea surface temperature relative to their driving global models, and all of them showed clear error reduction along coastlines and in semi-enclosed areas such as the Adriatic, Alboran, and Aegean Seas, where fine-scale ocean-atmosphere interactions and topographic constraints dominate. Yet when the team averaged errors across the entire basin, most regional models failed to correct biases in the mean, the standard deviation, the 90th percentile, and the long-term warming trend.

This failure at the basin scale is not a surprise to downscaling specialists, but its clarity here is striking. Large-scale biases in sea surface temperature are inherited from the driving global models, and no amount of regional refinement can remove them, because the regional model receives its boundary conditions from the very model whose errors it is meant to fix. The study found that regional models tended to amplify the cold bias present in many of their parent models rather than reduce it, and most also underestimated the observed warming trend. Trend errors showed no systematic spatial structure, deteriorating even in the coastal zones where other metrics improved, which points to a large-scale origin—possibly linked to the same aerosol-related radiation biases that have plagued European climate projections in other model families.

The picture brightens considerably when the analysis turns to the shape of the temperature distribution itself. Six of the nine regional models showed positive added value for median-removed sea surface temperature anomalies, meaning they captured the full statistical texture of day-to-day variability better than their parents, with improvements concentrated in the central and eastern Mediterranean, including the Tyrrhenian Sea. For the warmest extremes—temperatures above the 90th percentile—regional models generally stretched the upper tail of the distribution toward observations, bringing simulated heat extremes closer to what satellites actually record. One intriguing exception was the Gulf of Lion, where regional models produced narrower temperature distributions than their parents. The authors argue this may actually reflect a more physically consistent representation of deep winter convection, in which surface temperatures are anchored to the temperature of deep water layers, rather than a genuine deterioration in performance.

Marine heatwave duration emerged as the clearest success story of the entire study. Global models substantially overestimate how long heatwave events persist in the Mediterranean, and almost every regional model in the ensemble sharply reduced this bias, with improvements that were spatially widespread and robust across the Adriatic Sea in particular. The mechanism is physically coherent: regional models resolve the mesoscale wind systems and sharper sea surface temperature gradients that trigger rapid surface cooling, and it is precisely these atmospheric forcing episodes that terminate heatwave events. The added value showed up most strongly in the representation of short-lived events, which global models tend to miss entirely while simultaneously exaggerating the longevity of long-lived ones. Seven of nine models achieved positive basin-average added value for duration, with degradation confined to less than 8 percent of the basin.

Heatwave intensity told a different and more sobering story. Global models generally underestimate how hot marine heatwaves become, and while the majority of regional models improved this metric, three of them—AWI, LMD-CNRM, and GUF—showed clear deterioration, with degradation affecting 35 to 44 percent of the Mediterranean domain. The study traced these failures to model-specific configuration choices rather than resolution alone. The AWI model, for instance, uses a surface ocean layer 16 meters thick, thicker even than its parent global model, which damps short-term temperature fluctuations and weakens warm anomalies. Compelling evidence for this interpretation comes from the UNIBEL model: despite coarser horizontal resolution, its very thin 1.8-meter surface layer allowed it to achieve positive added value for both duration and intensity. When the team correlated added value against configuration factors across the ensemble, only one factor showed statistically significant relationships with both duration and intensity: the thickness of the uppermost ocean layer.

This finding carries a practical warning for the next generation of climate models. Higher horizontal resolution, the study concludes, is a necessary but not sufficient condition for accurately representing marine heatwaves. Simultaneous advances in vertical discretization, air-sea flux parameterizations, and atmospheric physics are equally essential, and the quality of the parent global model still matters enormously—regional models driven by badly biased parents showed the largest duration improvements, but intensity performance depended on how faithfully variability was transmitted across scales. The authors also note a methodological subtlety: because marine heatwaves are defined relative to each model’s own climatology, mean-state biases are implicitly removed, meaning the analysis captures only the variability component of added value. Future work using absolute thresholds could recover the mean-state dimension.

For the Mediterranean, a recognized marine heatwave hotspot where events are projected to grow more frequent, severe, and extensive, the stakes of getting these models right are high. Reliable regional projections of thermal stress persistence are critical for ecosystem managers, aquaculture operators, and coastal communities preparing for a hotter future. This study, the first systematic multi-model added value assessment for oceanic variables based on coordinated global-regional model pairs in any ocean region worldwide, provides both reassurance and a roadmap: regional downscaling genuinely improves the patterns, distributions, and duration of marine heatwaves, but the intensity of the most dangerous events will remain poorly constrained until model developers give as much attention to the top few meters of the ocean as they do to the grid that covers it.

Subject of Research: Added value of high-resolution coupled regional climate models for simulating Mediterranean sea surface temperature and marine heatwaves

Article Title: The added value of Med-CORDEX coupled high-resolution regional climate models in representing sea surface temperature and marine heatwaves in the Mediterranean Sea

Article References: De Rovere, F., Bonino, G., McAdam, R., Scoccimarro, E., Somot, S., Parras-Berrocal, I. M., Ahrens, B., Djurdjevic, V., Li, L., & Masina, S. (2026). The added value of Med-CORDEX coupled high-resolution regional climate models in representing sea surface temperature and marine heatwaves in the Mediterranean Sea. Ocean Science, 22(5), 2725-2742. https://doi.org/10.5194/os-22-2725-2026

Image Credits: AI Generated

DOI: 10.5194/os-22-2725-2026

Keywords: Mediterranean Sea, marine heatwaves, sea surface temperature, Med-CORDEX, regional climate modeling, downscaling, CMIP, ocean warming, climate extremes, air-sea interaction, ocean model resolution, climate projections

The stakes are far from academic. During 2011 and 2012, droughts across the central United States caused more than 30 million dollars in agricultural losses, and rapid-onset flash droughts, which intensify in a matter of weeks, have become an increasing concern for farmers, water managers, and emergency planners. Effective monitoring and detection tools are essential for managing these risks, yet most existing approaches to drought mapping share a fundamental weakness: they treat space and time as separate dimensions. A typical analysis first identifies spatial clusters of dry conditions at a given moment, then stitches those snapshots together across time. Because space and time are genuinely inseparable in a physical drought, that separation introduces error at every step, blurring the boundaries of events and sometimes merging distinct droughts into one another.

The new method, developed by T. Elana Christian and colleagues, tackles this problem head-on by treating space and time jointly at every stage of the analysis. The algorithm operates in two levels. The first level is a modified version of the classic k-means clustering algorithm, a workhorse of data science that partitions data points into a user-defined number of groups. In standard k-means, each data point is assigned to the cluster whose centroid, or center, is closest by some distance measure. When applied naively to a drought index, this produces clusters stratified purely by intensity, with locations flickering rapidly between drought states from one day to the next, a pattern the authors describe as nonphysical. Real landscapes do not transition between drought conditions instantly.

The innovation lies in how the modified algorithm defines closeness. Rather than measuring the distance between isolated space-time data points, the researchers first construct a spatiotemporal neighborhood around every point, consisting of adjacent locations and days whose values are correlated with the central point. The assignment step then averages distances across this neighborhood, so that a point’s cluster membership depends not just on its own intensity but on the intensity of its surrounding space-time context. The result is clustering based on both intensity and local spatiotemporal consistency. When the team compared the two approaches on data from 2003, the standard k-means produced erratic, rapidly shifting drought regimes, while the space-time version generated temporally persistent regimes with realistic durations and transitions that mirrored the physical evolution of drought.

Practical considerations shaped the implementation as well. The dataset is enormous: daily values of a drought index across the continental United States at a resolution of roughly 14 kilometers, spanning 42 years from 1980 to 2021. To handle data of this scale, the researchers employed a minibatched variant of k-means, which optimizes the clustering objective over random batches of data rather than the full dataset, trading a small amount of precision for dramatically faster convergence. The drought index itself, called the standardized vapor pressure deficit drought index, or SVDI, was introduced in 2022 as a simplified, temperature-based measure of atmospheric evaporative demand, calculated from daily maximum air temperature and minimum relative humidity. Higher SVDI values indicate drier conditions, and the index has previously proven capable of identifying known flash droughts and large-scale drought variability.

Between the two levels of the algorithm sits an optional, expert-informed step. After the space-time k-means clusters are formed, their mean values are computed and classified into drought severity categories, following conventions similar to those used for the standardized precipitation evapotranspiration index. Clusters whose means exceed 0.5, corresponding to mild, moderate, severe, or extreme drought, are selected and passed forward, while clusters in the same severity category are collapsed together. This intermediate filtering reduces the computational load on the second level and sharpens its performance. The second level then applies DBSCAN, a density-based clustering algorithm from 1996 that groups points lying in high-density regions and labels isolated points as noise. Crucially, DBSCAN does not require the number of clusters to be specified in advance and can carve out clusters of arbitrary shape, making it well suited to separating drought events that share similar statistical properties but occupy different regions of space and time.

The researchers tuned the full pipeline, which has six tunable parameters in total, against the well-documented summer 2003 flash drought in the central United States, an event marked by rapid intensification beginning in late June and a rapid demise before September. Spatiotemporal neighborhoods were sized according to the correlation structure of the data: temporal autocorrelation decayed within about a week, so a 14-day window was chosen, while spatial correlations became negligible beyond roughly 127 kilometers. The number of k-means clusters was set to five, and the DBSCAN neighborhood radius was selected by inspecting how the extracted clusters changed across a range of values. Too small a radius produced unrealistic, intensity-stratified clusters; too large a radius merged distinct events into one. The chosen value cleanly separated three distinct drought structures in 2003, including a large event spanning the southwestern and central United States, and the algorithm’s output aligned closely with the known spatial extent and timing of the historical drought.

A key test of robustness came when the team applied parameters tuned on 2003 to an entirely different year, 2000, which featured a flash drought originating in Florida, Mississippi, and Georgia alongside a longer-term drought across Texas, the Plains, and the Southwest. Even without retuning, the algorithm captured the southeastern flash drought with reasonable accuracy, briefly conflating it with the contemporaneous Texas and Plains drought for about two weeks before recovering the correct location. When the DBSCAN parameters were tuned specifically to 2000, the misidentification disappeared entirely, and the drought was classified as extreme rather than severe. The authors attribute the residual errors in the transfer test to the spatial smoothness of the drought index itself, which can link regions experiencing different events through continuously high values.

Running the algorithm year by year across the full 1980 to 2021 record revealed long-term shifts in American drought behavior. Mild and moderate droughts dominated the first half of the period, while severe and mild droughts predominated in the latter half. The frequency of mild droughts has increased markedly: between 1981 and 2000 there were years with no mild drought at all, but from 2000 to 2021 a mild drought occurred almost every year. Drought durations are changing too. After 2000, mild droughts became more likely to persist for an entire year, severe droughts are lasting longer and occurring more frequently than moderate ones, and the longest extreme drought in the dataset occurred in 2011. These trends correspond to known patterns in the drought literature, lending independent credibility to the method’s output.

Beyond drought, the framework is deliberately generalizable. Because the space-time width of the neighborhoods and the DBSCAN parameters can be adjusted, the method can be tuned to events at very different scales, from short-lived flash droughts to multi-season megadroughts, and the authors suggest it could be applied to other spatiotemporally continuous environmental phenomena. They also note that either level of the pipeline could be swapped for alternative algorithms, such as HDBSCAN, BIRCH, or Voronoi-based clustering, each offering advantages in speed or the ability to handle clusters of varying density. The code has been released publicly on GitHub, and the team highlights a remaining challenge common to unsupervised approaches: automating parameter tuning toward specific environmental phenomena, so that domain expertise, while still valuable, is no longer a prerequisite for turning raw climate data into a catalog of discrete, traceable events.

Subject of Research: A bi-level spatiotemporal clustering algorithm for extracting drought events from a vapor pressure deficit drought index over the continental United States

Article Title: A bi-level spatiotemporal clustering approach and its application to drought extraction

Article References: Christian, T. E., Subrahmanya, A. N., Gamelin, B., Rao, V., Samia, N. I., & Bessac, J. (2025). A bi-level spatiotemporal clustering approach and its application to drought extraction. Advances in Statistical Climatology, Meteorology and Oceanography, 11(2), 257-272. https://doi.org/10.5194/ascmo-11-257-2025

Image Credits: AI Generated

DOI: 10.5194/ascmo-11-257-2025

Keywords: drought, spatiotemporal clustering, k-means, DBSCAN, flash drought, vapor pressure deficit, SVDI, machine learning, climate extremes, United States, NLDAS, drought monitoring

The physical mechanism at the heart of the study is known as soil moisture–temperature coupling, or SM–T for short. Under the classic Budyko framework, high-latitude regions like Sweden sit firmly in the wet climate regime: evaporation there is limited not by how much water the soil holds but by how much energy the sun supplies. When soils are wet, incoming solar radiation is largely consumed by evaporating water, a process that cools the surface. But when enhanced warming drives soil moisture below a critical, non-stationary threshold, the system can shift into a transitional regime. There, rising temperatures deplete soil water, evaporation declines, and the energy that once went into vaporizing water is instead funneled into heating the air. This positive feedback can significantly magnify the intensity of an initial temperature extreme, turning a hot spell into a full-blown heatwave.

Sweden in 2018 offered a textbook case. The meteorological drought that began in late spring rapidly progressed into an agricultural drought as soil water reserves collapsed, and ultimately into a hydrological drought marked by reduced streamflow and groundwater storage. Such cascades are common in the mid-latitudes but highly unusual in the generally moisture-abundant high latitudes, which is precisely why the event attracted so much scientific attention. Previous work by Paul Dirmeyer and colleagues had already shown, using in situ and reanalysis data, that soil moisture–temperature coupling amplified the strength of the 2018 drought and heatwave in this very region. What remained unclear was exactly how long the coupling persisted, how much territory it covered, and which parts of the feedback chain mattered most.

To answer those questions, the team deployed WRF-CTSM, a recently developed coupling of the Weather Research and Forecasting atmospheric model with the Community Terrestrial Systems Model, the land surface component that corresponds to the Community Land Model version 5. Both components rank among the most sophisticated models in their respective domains. The researchers ran four simulations for 2018 at 10.5-kilometer horizontal resolution over a domain covering Norway, Sweden, and Finland, all initialized and nudged at the boundaries with the ERA5 reanalysis. The simulations differed in their land surface configurations: the reference run followed an earlier model setup, while subsequent runs successively incorporated a newer model version, the Medlyn stomatal conductance scheme, twenty hydrologically active soil layers reaching 8.6 meters deep, subgrid plant functional type heterogeneity, a plant hydraulic stress configuration that resolves water potential from roots to leaves, and finally a biomass heat storage scheme that accounts for the substantial thermal inertia of tree stems and leaves.

Benchmarking these simulations demanded independent evidence. The team drew on eddy-covariance flux tower observations from five ICOS stations within the study area, including Lanna, Norunda, Hyltemossa, Rosinedal-3, and Degerö, which provide daily measurements of surface soil moisture, sensible and latent heat fluxes, maximum two-meter temperature, precipitation, and net radiation. They supplemented these ground truth data with the European Space Agency’s Climate Change Initiative satellite soil moisture product and constructed a merged dataset called GLEAM-E-OBS, combining the Global Land Evaporation Amsterdam Model version 4.1a with the E-OBS gridded station-based product, both at a 0.1-degree grid spacing comparable to the model output. Crucially, the newest GLEAM release provides sensible heat flux for the first time, enabling the calculation of evaporative fraction, the ratio of latent heat to total turbulent energy exchange, which serves as a key indicator of land surface dryness.

The evaluation showed that the models and the merged dataset captured the magnitude and variability of the key coupling variables with impressive fidelity. Daily maximum temperature correlations against station observations reached 0.99 across all evaluated datasets. Surface soil moisture in the newer model runs aligned closely with both the towers and the satellite product, while the reference run ran noticeably dry, underestimating soil moisture by roughly 0.1 cubic centimeters per cubic centimeter at the station locations. Evaporative fraction proved the hardest variable to reproduce, as it fluctuates with intermittent cloudiness and local surface conditions. The newest configuration, incorporating both plant hydraulics and biomass heat storage, delivered the best overall performance, and it reduced summertime maximum temperatures over densely forested areas by up to 0.6 degrees Celsius by damping the diurnal temperature swing in ways that better match reality.

For the coupling analysis itself, the researchers devised a multi-correlation overlay approach built on 92-day running windows. Complete coupling was declared present only when three correlation conditions held simultaneously: a negative correlation between maximum temperature and surface soil moisture, a positive correlation between soil moisture and evaporative fraction, and a negative correlation between evaporative fraction and maximum temperature, all co-occurring with a rising temperature trend. This three-legged definition, adapted from Dirmeyer’s two-legged coupling metric, deliberately avoids reliance on critical soil moisture thresholds, which vary over time and are difficult to pin down. It also excludes cases where reduced evaporation stems purely from atmospheric dryness or plant stomatal closure without genuine soil moisture limitation, ensuring that only true transitions into the moisture-limited evaporation regime count.

The headline result is striking in its simplicity: averaged across the study region and all five datasets, complete soil moisture–temperature coupling persisted for 22 days during the May-to-August 2018 period. The individual legs of the feedback told a more nuanced story. The correlation between maximum temperature and soil moisture alone persisted for 72 days on average, and the soil moisture–evaporative fraction link for 42 days, but the atmospheric leg linking evaporative fraction to temperature lasted only 25 days. Yet that shortest-lived component turned out to be the gatekeeper: it coincided with the other two conditions 92 percent of the time, meaning that counting days in the atmospheric leg alone predicts the full coupling duration with high, though not definitive, probability. Geographically, the coupling concentrated in southeastern Sweden, where reduced evaporative fraction persisted longest and where some locations logged up to 30 consecutive heatwave days.

Differences among the datasets were revealing. GLEAM-E-OBS showed coupling over most of the region lasting 32 days on average, while the model simulations covered smaller areas and lasted 17 to 23 days, partly because the model overestimated precipitation during the dry period, injecting moisture pulses that shortened the feedback. Notably, unlike the merged observational product, WRF-CTSM correctly captured the anomalous character of 2018 relative to other years, which the authors attribute to its richer representation of vegetation physiology, including compensatory root water uptake from deeper soil layers during water stress, a process that field observations confirmed even in shallow-rooted spruce. Canopy transpiration emerged as the dominant component of total evaporation in this needleleaf-forest-dominated landscape, declining sharply in early July as the feedback intensified.

The implications reach well beyond one Scandinavian summer. Soil moisture–temperature interactions modify the surface energy balance in ways that can accelerate warming in Earth system models, so representing the coupling accurately directly affects projections of future heat extremes. As compound hot and dry events become more likely with global warming, and as high-latitude regions increasingly shift from energy-limited to moisture-limited evaporation during hot summers, understanding when and where the land amplifies the atmosphere becomes essential for seasonal forecasting, agricultural planning, and climate risk assessment. This study demonstrates that a coupled regional model with detailed plant hydraulics and heat storage physics can reproduce the anatomy of a land-atmosphere feedback with credible precision, while also highlighting how much local-scale timing and duration still elude even the best tools. The authors caution that their analysis covers a single event, and that applying the same framework to additional heatwaves will be needed to confirm the patterns, but the message is clear: the ground beneath a heatwave is not a passive victim. It is an active accomplice, and knowing when it joins the conspiracy may be one of the keys to anticipating the extremes of a warming world.

Subject of Research: Soil moisture–temperature coupling during the May–August 2018 heatwave in southern and central Sweden, analyzed with the WRF-CTSM coupled regional climate model and observational datasets

Article Title: Soil moisture–temperature coupling during extreme warm conditions in 2018 in Sweden: a case study with WRF-CTSM

Article References: Mužić, I., Hodnebrog, Ø., Yilmaz, Y. A., Berntsen, T. K., Sillmann, J., Lawrence, D. M., & Dirmeyer, P. A. (2025). Soil moisture–temperature coupling during extreme warm conditions in 2018 in Sweden: a case study with WRF-CTSM. Advances in Statistical Climatology, Meteorology and Oceanography, 11(2), 273-292. https://doi.org/10.5194/ascmo-11-273-2025

Image Credits: AI Generated

DOI: 10.5194/ascmo-11-273-2025

Keywords: soil moisture, temperature coupling, heatwave, Sweden 2018, land-atmosphere interactions, WRF-CTSM, evaporative fraction, drought, regional climate modeling, plant hydraulics, GLEAM, climate extremes

The innovation lies in what the researchers chose to work with. Widely used drought atlases, including the North American Drought Atlas, generally target a single variable: the summer Palmer Drought Severity Index, a measure of soil moisture that reflects a complex mixture of meteorological conditions. The index responds not only to rainfall but also to evapotranspiration, which depends on temperature, humidity, wind speed, and radiation, and it integrates conditions over the preceding year or more. That complexity, the team realized, is an opportunity rather than a limitation. Because the Palmer index is shaped by both temperature and precipitation, the reconstructed values published in the drought atlas can be mathematically inverted to infer the underlying meteorology, albeit with considerable uncertainty that must be handled rigorously.

To do this, the researchers built a hierarchical Bayesian model with three components. First, they estimated a forward model describing how the Palmer index responds to monthly temperature and precipitation, including a lag term capturing soil moisture memory, so that a dry year tends to be followed by another dry year. Second, they learned a prior distribution on the 24-dimensional vector of monthly temperatures and precipitation, spanning September through August, that captures the observed covariance structure between these variables. Third, they combined these elements to compute a full posterior distribution for unobserved past climate. The prior was constructed with an LKJ distribution on the covariance matrix, and the model was fitted using the PyMC probabilistic programming framework with the No-U-Turn Sampler, an adaptive variant of Hamiltonian Monte Carlo.

One unavoidable complication is that all modern weather occurs against the backdrop of rapid global warming. Because current global temperatures are unprecedented in the instrumental record, the basic assumption of most learning models, that training data resembles validation data, breaks down. The team addressed this by assuming that regional monthly temperatures and precipitation scale with global mean temperature, using the GISS Surface Temperature Analysis to estimate the anomaly. Over the reconstruction period, where no instrumental global temperature exists, they treated last-millennium global temperature as a random variable with a distribution informed by the variability of the detrended twentieth-century record. This allowed the model to separate the forced warming response from internal variability when learning what past climate might have looked like.

The three study locations were chosen deliberately, each a hotspot of past drought activity with a distinct precipitation climatology. Dodge City, Kansas, sits at the heart of the Dust Bowl, where rain falls mainly in spring and summer. Red Mesa, Arizona, near the Four Corners region, is the epicenter of historical megadrought activity, with precipitation spread more evenly through the year. Sonora Junction, California, in the Sierra Nevada, receives almost all its precipitation in winter. The fitted forward model revealed these differences clearly: in Kansas, high summer drought index values were robustly associated with cool, wet springs and summers; in Arizona, the index was more sensitive to fall and winter precipitation; and in California, winter rainfall and even small amounts of June rain strongly influenced the index, while the previous year’s conditions mattered little.

Validation against the CRU TS instrumental dataset showed the method performs well, especially for precipitation. Correlations between the posterior mean reconstruction and observations over the validation period reached 0.72 for Dodge City, 0.72 for Sonora Junction, and 0.78 for Red Mesa. The reconstruction captured the Dust Bowl drought in Kansas better than the multi-proxy Last Millennium Reanalysis, and instrumental precipitation values fell within the reconstructed likely range for most years. Temperature proved harder: correlations of 0.54, 0.31, and 0.4 across the three sites reflected the simple fact that the growing-season Palmer index is far more sensitive to rainfall than to temperature. The method still showed some skill, shifting posterior temperatures toward higher values during the scorching 1930s, even if it could not capture the full magnitude of the extreme heat.

The probabilistic framework also enables something deterministic reconstructions cannot: statements about the odds that particular years or decades were the most extreme in a thousand years. By generating 4000 realizations of the posterior and concatenating them with instrumental data, the team calculated the probability that any given year was the wettest or driest on record. In Kansas, the year 1522 was the most likely driest year of the millennium, with a 13 percent probability, while 1913, the driest year in the instrumental record, had a 23 percent chance of being the driest since at least 1000. In Arizona, there was only a 0.05 percent chance the driest year occurred after 1901, with candidates including 1748, 1685, 1729, and 1542. In California, the driest year most likely fell in 1580, 1053, 1729, or 1126.

The decade-scale results carry the most dramatic implications. There is a 71 percent probability that 1931 to 1940 was the driest decade on record in central Kansas, confirming the Dust Bowl as a hydroclimatic event essentially without precedent in the past millennium. Meanwhile, the late twentieth century stands out on the wet end of the distribution: a 15 percent chance that 1992 to 2001 was the wettest decade in Kansas, and a 31 percent chance that 1979 to 1988 was the wettest decade in the Arizona region. These findings reinforce recent work identifying an extraordinary late twentieth-century pluvial in the American southwest, a wet episode that may have masked the onset of the aridification now underway.

The study also exposes an important limitation with implications for climate risk assessment. Over the training period of 1950 to 2005, the Kansas region simply did not experience hot droughts of the kind that struck in 1934, when record-shattering heat accompanied extreme precipitation deficits. The 1956 drought brought similarly low summer rainfall, but average summer temperatures were 1.67 degrees Celsius colder than in 1934. Because tree-ring drought records from the late twentieth century are relatively insensitive to temperature, a model trained on that era will systematically fail to capture combined heat and drought extremes, precisely the compound events that pose the greatest dangers in a warming world.

The authors emphasize that their method is designed to complement, not replace, existing reconstruction techniques that assimilate multiple proxies or rely on climate model priors. Because it is fully Bayesian, the framework can be updated as new data arrive and extended to incorporate greater complexity, such as autoregressive or Gaussian process models of internal variability, or independent reconstructions of global mean temperature. For now, the study demonstrates how much hidden information can be coaxed from a single, familiar dataset. The rings of centuries-old trees, it turns out, were quietly recording not just drought, but the temperature and rainfall combinations that produced it, waiting for the right statistical tools to set them free.

Subject of Research: Bayesian reconstruction of last-millennium temperature and precipitation from tree-ring-based drought atlas data

Article Title: Joint probabilistic estimates of temperature and precipitation from tree ring-based reconstructions of the last millennium

Article References: Marvel, K., Cook, B., Weng, E., Singh, R., & Cook, E. (2026). Joint probabilistic estimates of temperature and precipitation from tree ring-based reconstructions of the last millennium. Advances in Statistical Climatology, Meteorology and Oceanography, 12(1), 43-57. https://doi.org/10.5194/ascmo-12-43-2026

Image Credits: AI Generated

DOI: 10.5194/ascmo-12-43-2026

Keywords: tree rings, paleoclimate, Bayesian inference, drought, Palmer Drought Severity Index, North American Drought Atlas, Dust Bowl, precipitation reconstruction, temperature reconstruction, last millennium, megadrought, climate variability

The core idea behind downscaling is deceptively simple. Global climate models divide the planet into grid cells that are typically one hundred kilometres or more across, a resolution at which coastlines, mountain ranges, and individual thunderstorms are essentially invisible. Regional climate models take the large-scale output from these global models and re-simulate it over a smaller domain at much finer resolution, adding back the topographic and coastal complexity that shapes local weather. In Australasia, where the Great Dividing Range, the complex coastline of Bass Strait, and the steep rainfall gradients of the tropical north all exert powerful influences on climate, this added detail is not a luxury but a necessity for credible projections.

The CORDEX-Australasia programme has now completed successive ensembles based on the fifth and sixth phases of the Coupled Model Intercomparison Project, known respectively as CORDEX-CMIP5 and CORDEX-CMIP6. These coordinated collections of regional simulations, produced by modelling groups across Australia and beyond, allow scientists to separate the robust signals of climate change from the noise of model-to-model differences. The synthesis of this fifteen-year effort, published in PLOS Climate by Jason Evans of the University of New South Wales and a large team of collaborators from institutions across the region, documents both the scientific achievements and the formidable challenges that remain.

One of the clearest findings to emerge from the programme is that regional models demonstrate genuine added value over the global models that drive them. This is not merely a matter of sharper pictures. The high-resolution simulations represent temperature extremes, heavy precipitation events, coastal processes, and complex terrain with substantially greater fidelity, which matters because these are precisely the phenomena that cause the most damage to communities and infrastructure. When a regional model can resolve the sea breeze that moderates a coastal heatwave, or the orographic uplift that triggers a flash flood on a mountain slope, its projections of future extremes become far more useful to planners and emergency managers.

The CMIP6-based projections now available for Australasia show robust agreement on one central fact: substantial warming through the twenty-first century is essentially locked in across the region. The magnitude of that warming, however, depends overwhelmingly on which emissions pathway the world follows, making human choices the dominant source of uncertainty in temperature projections. Precipitation is a different story altogether. Rainfall responses vary considerably from one subregion to another, with some areas projected to become drier and others facing changes that are harder to pin down, reflecting the intricate interplay of atmospheric circulation, ocean temperatures, and local geography that governs water availability across the continent and its island neighbours.

Beyond the headline projections, the past decade and a half has seen significant methodological advances that have quietly revolutionised the field. Convection-permitting modelling, which runs regional models at kilometre-scale resolutions fine enough to simulate thunderstorms directly rather than approximating them with statistical formulas, has moved from research curiosity to operational tool. Benchmarking frameworks now allow systematic evaluation of model performance against observations, while improved bias-correction methods help translate raw model output into information that hydrologists, engineers, and farmers can actually use. Together these advances have sharpened both the physical realism of the simulations and the credibility of their evaluation.

The practical payoff of this scientific investment is now visible across nearly every sector of Australian and Pacific society. Regional climate projections underpin decision-making in water resource management, where reservoir operators need to know how inflows will shift in the decades ahead. Agricultural planners use the projections to assess how growing seasons and crop water requirements will change. Disaster risk reduction agencies, energy system planners, public health authorities, and national climate services all draw on the same coordinated datasets, meaning that a single scientific infrastructure quietly supports adaptation decisions worth billions of dollars across the region.

None of this would have been possible without sustained investment in shared computing infrastructure. The National Computational Infrastructure in Australia has provided the enormous processing power and data storage that regional climate modelling demands, and coordination has been strengthened further through the National Partnership for Climate Projections, which links state and federal programmes into a coherent whole. This model of collaboration, in which different levels of government and multiple research institutions contribute to a common international framework, has become a quiet success story in how complex scientific infrastructure can be built and maintained over decades.

Significant challenges nonetheless remain on the horizon. Regional climate modelling is computationally voracious, and the demand for ever-finer resolution continually pushes against available supercomputing capacity. Persistent model biases, particularly in simulating precipitation, continue to limit confidence in some projections. Designing ensembles that efficiently sample the range of possible futures without duplicating effort is an ongoing scientific puzzle, and perhaps the most difficult challenge of all is communicative rather than computational: translating sprawling multi-model datasets into decision-ready guidance that a council engineer or a hospital administrator can act upon without needing a doctorate in climate science.

The strategic priorities identified for the coming years point toward an ambitious future. Kilometre-scale convection-permitting modelling is expected to become the new standard, capturing storm-scale processes that current models still approximate. Enhanced coupling between regional models and Earth system components such as oceans, vegetation, and carbon cycles promises more complete simulations. Coordinated super-ensembles combining many models and scenarios, the responsible integration of artificial intelligence techniques to accelerate both simulation and analysis, stronger observational constraints to narrow uncertainty, and expanded support for vulnerable remote Australian islands and Pacific Island nations all feature prominently in the roadmap. For the low-lying atoll nations of the Pacific, whose very existence is threatened by rising seas and intensifying storms, the stakes of this work could hardly be higher.

The fifteen-year story of CORDEX-Australasia is ultimately a story about turning planetary-scale physics into locally meaningful knowledge. As the synthesis makes clear, delivering robust, high-resolution climate information for adaptation and resilience across Australasia will require sustained investment, continued collaboration, and relentless innovation. But the foundation has been laid, the datasets exist, the methods have matured, and the region now possesses a scientific capability that its founders could scarcely have imagined when the first coordinated downscaling experiments began. The task ahead is to ensure that this capability keeps pace with a changing climate and the growing urgency of the decisions it must inform.

Subject of Research: Regional climate downscaling and high-resolution climate projections for Australasia under the CORDEX programme

Article Title: Fifteen years of regional climate downscaling in CORDEX-Australasia

Article References: Evans, J. P., Thatcher, M., Argüeso, D., Di Luca, A., Trancoso, R., Andrys, J., Chapman, S., Di Virgilio, G., Dowdy, A., Fita, L., Gibson, P. B., Grose, M., Howard, E., Ji, F., Kala, J., Katzfey, J., Khadke, L., Mughal, O., Ng, B., … Ye, H. (2026). Fifteen years of regional climate downscaling in CORDEX-Australasia. PLOS Climate, 5(9), e0001066. https://doi.org/10.1371/journal.pclm.0001066

Image Credits: AI Generated

DOI: 10.1371/journal.pclm.0001066

Keywords: CORDEX-Australasia, regional climate modelling, climate downscaling, CMIP6, convection-permitting models, climate extremes, precipitation projections, Australasia, Pacific Islands, climate adaptation, bias correction, climate services