The research, led by Sina Dadras and Ehsan Hajighasemi of the University of Tehran’s Department of Environmental Engineering together with colleagues Khosro Ashrafi and Majid Shafiepour Motlagh, introduces a framework the authors call the Residual-Aware Meta Ensemble, or RAME. The central problem the team set out to solve is a familiar one in air quality modeling: conventional forecasting systems tend to perform reasonably well on average but systematically underestimate the extreme values that matter most. When ozone climbs toward a health-threatening peak during the photochemically active hours around midday, the difference between a forecast of 90 parts per billion and an actual reading of 120 parts per billion is not a statistical footnote. It is the difference between an ordinary afternoon and a day on which children, the elderly, and people with respiratory disease should stay indoors.

The architecture of RAME reflects a two-stage philosophy that has deep roots in machine learning but has rarely been tuned so explicitly for peak capture. In the first stage, an ensemble of base learners generates foundational forecasts. Each base learner is trained on the previous 24 hours of historical pollutant concentrations combined with real-time meteorological parameters such as temperature, solar radiation, wind speed, and humidity, all of which govern the photochemical reactions that produce and destroy ozone near the surface. The ensemble approach matters because individual models carry different inductive biases: a recurrent network may track the diurnal rhythm of ozone well, while a convolutional component may excel at detecting abrupt changes. By weighting and combining their outputs, the ensemble smooths out the idiosyncratic errors of any single learner, a strategy the authors trace back to the stacked generalization framework introduced by David Wolpert in 1992 and refined by Leo Breiman’s work on stacked regressions.

The second stage is where RAME earns its name. Rather than accepting the weighted ensemble forecast as the final answer, the framework trains a meta-learner to model the residual, meaning the discrepancy between what the ensemble predicted and what the monitoring stations actually measured. This residual-correction mechanism is built from bidirectional recurrent units paired with dilated convolutions. The bidirectional design allows the network to read the time series both forward and backward, capturing how an ozone peak is shaped by the hours that precede it and the conditions that follow it. Dilated convolutions, which skip inputs at increasing intervals, let the model perceive patterns across multiple time scales simultaneously, from the sharp hour-by-hour fluctuations driven by traffic cycles to the slower multi-day evolution of heat waves and stagnant air masses that set the stage for severe episodes.

Trained and evaluated on observations from eight air quality monitoring stations distributed across Tehran, RAME delivered a mean 24-hour averaged R-squared of 0.87, with a root mean square error of 8.12 parts per billion and a mean absolute error of 5.30 parts per billion. Those figures represent the model’s ability to explain nearly nine-tenths of the variance in hourly ozone concentrations across a city whose monitoring sites span dramatically different microenvironments, from dense traffic corridors to residential districts shaped by the mountains that wall in the Tehran basin. Crucially, the model consistently outperformed every individual baseline it was compared against, and it did so across heterogeneous urban sites rather than excelling in one favorable location while failing elsewhere. For a forecasting system intended to guide city-wide public health decisions, that spatial consistency is arguably as important as the headline accuracy numbers.

The most striking result concerns the events that regulators and health officials care about most. RAME correctly identified 96 percent of critical ozone days, the episodes on which concentrations approached or exceeded regulatory thresholds. The authors attribute this skill to the residual-correction stage, which learns precisely where and when the ensemble’s smoothed predictions fall short. Conventional systems, the study notes, tend to significantly underestimate concentrations during photochemical peak periods, because the statistical averaging that makes a model robust on typical days also flattens its response to extremes. By explicitly training a second network on the ensemble’s failures, RAME converts the systematic bias of the first stage into a learnable signal, effectively teaching the system to recognize the atmospheric fingerprints of an impending exceedance.

The choice of Tehran as the test bed is not incidental. The city sits at the southern edge of the Alborz mountains, in a semi-enclosed basin where temperature inversions, intense summer solar radiation, and a large vehicle fleet combine to create conditions that are almost laboratory-perfect for photochemical ozone production. Previous research by the same group and collaborators has documented the substantial health burden and economic cost of ambient air pollution in Tehran, and studies of summertime ozone and nitrogen oxide dynamics between 2017 and 2019 have mapped the regional contributions to the city’s ozone problem. Against that backdrop, an hourly forecasting tool that reliably flags peak events has immediate operational value: it can inform traffic restriction schemes, industrial curtailments, school activity decisions, and public health advisories issued hours before the worst air arrives rather than after.

The methodological lineage of the work is also worth noting. Over the past decade, researchers have applied long short-term memory networks, gated recurrent units, convolutional architectures, temporal convolutional networks, and transformer models to pollutant forecasting in cities from Beijing to Hangzhou to Bogota. Ensemble and hybrid approaches have repeatedly shown advantages over single models, and national-scale exposure mapping efforts in China and the United States have demonstrated the scalability of machine learning for ozone estimation. What distinguishes RAME is its explicit focus on the residual structure of ensemble error at hourly resolution, and its demonstration that a meta-learner can be trained to compensate for the specific failure mode, peak underestimation, that has limited earlier systems. The authors describe the framework as robust, scalable, and methodologically adaptable, suggesting that the same residual-aware design could be transferred to other pollutants and other cities without fundamental redesign.

Limitations and open questions remain, as they do in any modeling study. The framework depends on the quality and continuity of monitoring data, and missing observations in air quality networks remain a persistent challenge that researchers have addressed with imputation techniques ranging from interpolation to nearest-neighbor methods. The model’s reliance on 24 hours of historical data means its skill is anchored to the recent past, and rapid synoptic changes or unprecedented emission events could still surprise it. The authors also note that data will be made available on request, which opens the door to independent validation by other research groups. Nevertheless, the performance achieved across eight heterogeneous stations in one of the world’s most meteorologically challenging megacities suggests that the approach is more than a local curiosity.

The broader significance of the study lies in what it says about the future of urban air quality management. As climate change intensifies heat waves and stagnation episodes across the Middle East and beyond, the photochemical conditions that generate dangerous ozone are expected to become more frequent and more severe. Cities cannot remove ozone from the atmosphere the way they can remove a point source of particulate matter; they can only anticipate it and act. A forecasting system that captures 96 percent of critical ozone days, with errors measured in single-digit parts per billion, converts air quality management from a reactive exercise into a genuinely predictive one. For Tehran’s residents, and potentially for millions of others living in sun-baked basins and valleys around the world, that shift could mean the difference between breathing a warning and breathing the peak itself.

Subject of Research: Hybrid deep learning ensemble forecasting of hourly surface ozone concentrations in Tehran, Iran

Article Title: Hourly ozone prediction using ensemble forecasting to accurately capture peak concentration: a case study of Tehran, Iran

Article References: Dadras, S., Hajighasemi, E., Ashrafi, K., & Motlagh, M. S. (2026). Hourly ozone prediction using ensemble forecasting to accurately capture peak concentration: a case study of Tehran, Iran. Environmental Science and Pollution Research, 33(28), 14217-14240. https://doi.org/10.1007/s11356-026-38197-7

Image Credits: AI Generated

DOI: 10.1007/s11356-026-38197-7

Keywords: ozone prediction, ensemble forecasting, deep learning, Tehran, air quality, peak concentration, residual correction, meta-learner, urban pollution, public health, machine learning, Environmental Science and Pollution Research

The challenge stems from the nature of wildfire emissions themselves. Unlike the steady biogenic emissions from forests or the well-documented output of industrial sources, wildfires are rare, explosive and deeply unpredictable events. The location, size and intensity of new fires can only be estimated probabilistically, and the height at which smoke is injected into the atmosphere remains highly uncertain. Because air quality forecasts depend so heavily on emissions, these uncertainties cascade through the entire dispersion calculation, producing errors that vary enormously in space and time. The 2023 Canadian fire season, which consumed more than 15 million hectares of forest, made these limitations painfully visible as smoke crossed the Atlantic and degraded air quality in European cities.

Operational air quality assimilation systems, which blend model forecasts with surface observations to produce analyses, rely on accurate estimates of background forecast errors. Yet the error formulations used in most operational systems are highly simplified, often assuming uniform statistical properties that cannot capture the anisotropic, case-dependent error structure of a wildfire plume. Ensemble forecasting, the gold standard in numerical weather prediction, offers a way around this, but it comes at a steep price. Building a meaningful ensemble for air quality requires perturbing emissions, deposition rates and boundary conditions whose error characteristics are poorly known, and even ensembles of 50 to 100 members are orders of magnitude too small to sample the enormous dimensionality of the atmospheric system, leading to sampling errors that demand additional fixes like localization and inflation.

Vogel and her colleagues propose a fundamentally different approach rooted in ideas that trace back to stochastic dynamic prediction in the 1960s. Instead of running dozens of parallel forecasts, their method, called a parametric uncertainty forecast, evolves the error statistics themselves as prognostic variables within the forecast model. The key insight is that for a Gaussian error distribution, a handful of parameters, principally the error standard deviation and its spatial correlation length, are enough to describe the leading structure of the uncertainty. By deriving explicit equations for how these parameters change under advection, diffusion and emissions, the researchers can propagate uncertainty through the model at a tiny fraction of the computational cost of an ensemble.

The theoretical core of the paper lies in deriving these prognostic equations for the three processes that matter most for near-surface smoke forecasting. Advection turns out to be the simplest case: error standard deviation is transported by exactly the same wind field as the concentration itself, so the model’s existing advection scheme can handle it without modification. Vertical diffusion is more subtle. When a spatially correlated concentration field diffuses, its errors shrink, and the researchers show that this reduction depends on the vertical error correlation length, a new parameter that must be supplied to the model. Emissions act as a source of uncertainty, increasing concentration error in proportion to the correlated emission error standard deviation, a quantity that combines how uncertain the emissions are with how those errors relate to existing concentration errors.

A particularly important theoretical result concerns the choice between error variance and error standard deviation as the prognostic variable. Although the two are mathematically equivalent, the standard deviation form yields simpler equations with fewer terms, avoids an extra spatial derivative that introduces discretization errors, and shares the same units as concentration itself, making implementation and interpretation far easier. In the diffusion equation, the variance form requires an additional gradient term that the standard deviation form does not, and idealized experiments showed the standard deviation formulation tracks a full Kalman filter more accurately. The team therefore selected standard deviation as the working variable throughout their implementation.

The researchers embedded this parametric uncertainty forecast into GEM-MACH, Canada’s operational coupled meteorology and air quality model, which runs at 10-kilometer resolution over North America with 84 vertical layers and hourly wildfire emissions from the Canadian Forest Fire Emissions Prediction System. They chose fine particulate matter, PM2.5, as the target species and added its error standard deviation as a new prognostic field. For initial and boundary conditions, they replaced the operational bounded statistical fit between concentration and error variance with a simple linear relation applied to the instantaneous forecast concentrations, ensuring that uncertainties from emissions inside the simulation period remain consistent with those entering from outside it.

To demonstrate the approach, the team applied it to the June 6, 2023 smoke episode, when large fires in central Quebec sent plumes toward populated areas while older smoke lingered over Ottawa and Montreal. The parametric forecast produced dramatically different error fields from the operational setup: large, plume-shaped uncertainty structures grew directly from the remote fire hotspots and were carried southward by the winds, while the operational error fields, derived from two-month running statistics, remained nearly flat across the smoke regions. When these error estimates were fed into the hourly surface analysis system, the differences were striking. In well-observed urban areas, the two experiments produced similar corrections, but in the sparsely monitored north, the parametric experiment generated analysis increments exceeding 80 micrograms per cubic meter in plumes located roughly 400 kilometers from the nearest observation stations.

That last result may be the most consequential. Because the analysis spreads observation information according to the background error covariance, and because the parametric forecast assigns very large error standard deviations within the plumes, even weakly correlated distant observations can now exert meaningful corrections on unobserved smoke. The researchers emphasize that these far-reaching increments arise purely from the improved variance fields, not from exaggerated error correlations, meaning the sparse observing network is used far more efficiently in a highly case-dependent and anisotropic way. The method can even produce opposite-signed corrections within a single plume when nearby observations disagree, something the operational setup cannot achieve.

The authors are careful to note that their formulation remains deliberately simplified. Emission errors are assumed proportional to local emissions, the vertical error correlation length is prescribed rather than forecast, and processes like area emissions, chemistry, deposition and wind uncertainty are excluded for now. Sensitivity tests showed the results depend on the chosen correlation length, with shorter lengths producing shorter uncertainty plumes. Still, the study marks the first implementation of parametric diffusion and emission error dynamics in a state-of-the-science atmospheric model and leads the way toward a full parametric Kalman filter for operational smoke assimilation. As fire seasons intensify and public demand for trustworthy air quality warnings grows, this efficient new mathematics may soon help forecasters tell not just where the smoke will go, but how sure they are when they say so.

Subject of Research: Parametric uncertainty forecasting for wildfire smoke data assimilation in operational air quality models

Article Title: Formulation of parametric uncertainty forecasts towards operational wildfire smoke assimilation

Article References: Vogel, A., Ménard, R., Abu, J., & Chen, J. (2026). Formulation of parametric uncertainty forecasts towards operational wildfire smoke assimilation. Nonlinear Processes in Geophysics, 33(3), 347-371. https://doi.org/10.5194/npg-33-347-2026

Image Credits: AI Generated

DOI: 10.5194/npg-33-347-2026

Keywords: wildfire smoke, air quality forecasting, data assimilation, parametric Kalman filter, uncertainty quantification, GEM-MACH, PM2.5, error covariance, ensemble forecasting, emissions uncertainty, Canada wildfires, atmospheric modeling

The logic behind the approach starts with an inventory of what each information source can and cannot deliver. Non-initialized historical simulations, such as the 163 CMIP6 transient runs the team used, capture the forced response to greenhouse gases and aerosols but scatter widely around it because each member carries its own random realization of internal variability. Initialized decadal predictions, of which the study used 92 hindcasts from the CMIP6 Decadal Climate Prediction Project, are nudged toward the observed state of the ocean and atmosphere at their start date, which should in principle phase-lock the simulated modes of variability with the real ones. In practice, however, their skill often fades after a few years, and they suffer from initialization shock — a drift caused by the mismatch between biased models and the observed states they are fed. Observations alone can also constrain projections, as earlier work by Bonnet and others showed when selecting ensemble members using observed proxies of the Atlantic meridional overturning circulation. What had not been tried until now was blending all three sources in a single, computationally cheap framework.

The method, which the authors call BLEND, works as a two-stage member selection from the pool of historical simulations. In the first stage, each of the 163 historical runs is scored against an observational metric over a calibration window — typically twenty years — preceding the forecast start date. The similarity is quantified by combining normalized root-mean-square error and correlation, and the fifty best-matching simulations are retained. In the second stage, those fifty candidates are further filtered down to thirty by checking which ones come closest to the ensemble-mean temperature that the initialized decadal hindcasts actually predict for the forecast window over the region of interest. The result is a constrained subset of projections that agrees both with what the real climate has recently done and with what the prediction systems say it is about to do. Because the selection operates on non-initialized simulations, the method sidesteps the drift problem that plagues decadal forecasts by construction, while still harvesting their information.

A crucial design question is which observational metric to use as the first-stage constraint. The team tested four candidate indices, each chosen for its documented influence on European climate at annual to multidecadal timescales. The Atlantic Multidecadal Variability index captures basin-wide North Atlantic sea surface temperature fluctuations after removing the externally forced signal, using a ten-year low-pass Lanczos filter. The subpolar gyre index tracks SSTs over the 50–60° N, 15–40° W region, a hotspot where decadal predictability is known to be high. A global index measures nine-year pattern correlations between simulated and observed sea surface temperature fields worldwide, following earlier constraining studies. Finally, the winter North Atlantic Oscillation index, computed as the pressure difference between broad southern and northern boxes over the Atlantic, probes the dominant mode of atmospheric circulation variability. Observational references came from the NOAA ERSSTv5 dataset for the ocean indices and the ERA5 reanalysis for the NAO.

To test whether the blending actually helps, the researchers ran a retrospective evaluation every year from 1966 to 2000, producing five-, ten-, and fifteen-year forecasts of summer and winter surface temperature averaged over three IPCC reference regions: Northern Europe, West Central Europe, and the Mediterranean. They compared spread, defined as the range between the minimum and maximum ensemble members, and absolute error of the ensemble mean against the ERA5 observations. Three skill scores anchored the assessment: the mean squared skill score for deterministic performance, and the ranked probability skill score and continuous ranked probability skill score for probabilistic skill, all benchmarked against the full unconstrained historical ensemble. Critically, the team used a resampling test — drawing one thousand random thirty-member subsets from the historical pool — to verify that any improvement was not just an artifact of working with a smaller ensemble. A result counted as significant only when it beat the 95th percentile of those random draws.

The headline finding is that the constrained forecasts significantly narrow the uncertainty range. Nearly all observation-constrained subsets showed a pronounced spread reduction relative to both the full historical ensemble and the decadal hindcasts, across regions, seasons, and lead times, and the double-constrained BLEND subsets reduced spread even further, particularly at the five-year horizon. In an illustrative case, a five-year summer temperature forecast for West Central Europe starting in 1983 saw the historical ensemble predict −0.09 ± 0.81 °C while the blended forecast landed at −0.23 °C with a tighter spread and an ensemble mean closer to the observed value. The decadal hindcasts, by contrast, showed spread and error statistics broadly similar to the unconstrained projections, a disappointing result the authors attribute to possible initialization shocks, structural model biases such as a southward-shifted NAO pattern, and the fact that European decadal temperature variability is intrinsically weak compared to chaotic atmospheric noise.

Regional and seasonal texture matters enormously in the results. Over West Central Europe in summer, both the observation-only and blended forecasts using regional surface temperature as the predictor delivered large, statistically significant improvements at all three forecast horizons, and the blended versions generally outscored their observation-only counterparts — direct evidence that adding decadal prediction information pays off there. A spatial case study showed the advantage extending into Germany and Poland for ten-year summer forecasts. Over Northern Europe, significant gains appeared for five- and ten-year summer forecasts, and blended forecasts using the global SST pattern index improved fifteen-year predictions. In the Mediterranean in winter, the Atlantic Multidecadal Variability-based blends captured part of the low-frequency variability, consistent with the known role of the AMV in modulating the region’s decadal climate, though the regional gains were more modest and concentrated over parts of Spain and the Maghreb. Notably, the second, region-specific selection step could improve skill over the targeted region while degrading it elsewhere, a trade-off the authors flag for anyone applying the method.

The study is equally candid about failures and caveats. The NAO-based constraint performed poorly over Northern Europe despite the well-established physical teleconnection, possibly because CMIP-class models misplace the NAO’s centers of action or underestimate its decadal variability, or because a twenty-year calibration window is too long for an index with short decorrelation times. Over the Mediterranean in summer, where anthropogenic forcing dominates, none of the blended subsets beat random ensemble subsets in absolute error. The authors also stress that conclusions can flip depending on which skill score is used: improvements in the ensemble mean sometimes vanished once the full probability distribution was penalized for overconfidence, which is why they recommend deploying multiple complementary metrics aligned with the end user’s decision context. The optimal predictor, they conclude, varies by region and must be evaluated case by case.

The broader significance lies in what the framework makes possible. It runs at very low computational cost, requires no new simulations, and produces seamless climate information that stays consistent from the historical record through the coming decades — exactly the property adaptation planners need when they cannot afford a discontinuity between what has been observed and what is forecast. The authors suggest natural extensions to climate extremes such as heatwaves and droughts, and to other regions where a variable exhibits genuine multi-annual to decadal internal variability with identifiable large-scale drivers. If those extensions succeed, the humble act of cherry-picking the right simulations — guided jointly by what the real world has done and what initialized models expect — could become a standard tool for turning the noisiest part of the climate forecast into actionable, regionally tailored information.

Subject of Research: Reducing near-term regional climate forecast uncertainty by constraining climate model ensembles with combined observational and decadal prediction information

Article Title: Reduction of uncertainty in near-term climate forecast by combining observations and decadal predictions

Article References: Bonnet, R., Boé, J., Sanchez-Gomez, E., & Cassou, C. (2026). Reduction of uncertainty in near-term climate forecast by combining observations and decadal predictions. Earth System Dynamics, 17(5), 1299-1314. https://doi.org/10.5194/esd-17-1299-2026

Image Credits: AI Generated

DOI: 10.5194/esd-17-1299-2026

Keywords: decadal climate prediction, internal climate variability, CMIP6, climate projections, member selection, Europe, Atlantic Multidecadal Variability, North Atlantic Oscillation, forecast uncertainty, adaptation planning, Earth System Dynamics, ensemble forecasting

The team, led by Martin Bonte with Lesley De Cruz, Fabian Debal and Stéphane Vannitsem, evaluated the two models over Belgium using the RADCLIM radar product, a quantitative precipitation estimate that merges radar measurements with rain gauge observations through Kriging with External Drift. Crucially, neither model was fine-tuned for the Belgian domain. LDCast was used with its original pre-trained weights, and STEPS was run through the open-source pysteps library exactly as a national weather service would deploy it. This choice matters, because in practice most forecast offices will not have the time, data or expertise to retrain a state-of-the-art AI model before using it. The evaluation therefore reflects a realistic operational scenario rather than a best-case laboratory demonstration.

The two models represent fundamentally different philosophies of nowcasting. STEPS, short for Short-Term Ensemble Prediction System, is built on Lagrangian persistence: it advects the observed rainfall field along a motion field estimated by optical flow, decomposes the field into a cascade of spatial scales, and lets each scale evolve through an auto-regressive process. Randomness enters through stochastic perturbations of both the motion field and the intensities. LDCast, by contrast, is a latent diffusion model. A variational autoencoder compresses sequences of rainfall fields into a compact latent space, a forecaster network predicts the latent representation of future fields, and a denoiser stack then generates different ensemble members conditioned on that output. The generative approach naturally produces diverse, nondeterministic forecasts, which makes ensemble construction almost effortless.

To compare the two fairly, the researchers selected ten convective and ten stratiform rainfall events, classified using ERA5 reanalysis data on convective and large-scale precipitation rates. For each event they generated 50-member ensembles and examined the spread-error relationship scale by scale, using Fourier power spectra. In a perfectly calibrated ensemble, the spread among members should match the error of the ensemble mean against observations. The spectral analysis showed that both models come remarkably close to this ideal for scales larger than about five kilometres. LDCast was slightly underdispersive overall, though somewhat overdispersive at large scales during convective events, while STEPS was underdispersive at short lead times for stratiform rain and became well calibrated as lead time grew.

One subtlety the authors had to untangle involved the smallest scales, below roughly five kilometres, where the power spectrum of radar observations flattens out. This flattening is not a physical feature of rainfall but an artefact of white noise contamination in the radar data and of the smoothing introduced by merging with rain gauges. The apparent underdispersion of both models at these scales therefore cannot be blamed on the ensembles themselves. Interestingly, the error at these smallest scales saturates almost immediately, within the first five minutes of the forecast, at scales between five and ten kilometres depending on the model and event type, meaning the issue is largely irrelevant for practical forecasting.

Beyond the raw size of uncertainty, the team probed the internal geometry of the ensembles using the covariance matrix of the members and its eigenvectors, essentially a principal component analysis of the forecast uncertainty. The results revealed a striking architectural difference. STEPS ensembles, especially in convective cases, are dominated by a handful of directions in phase space: convective ensembles concentrate their variability in roughly ten dominant modes, and stratiform ensembles in about five. The residual vectors of individual members align along these few leading eigenvectors, meaning the members are quite similar to one another. LDCast ensembles, in contrast, display much more homogeneous eigenvalue spectra, spreading their variability more evenly across many modes. Both models, however, adapt their ensembles to the meteorological situation, developing larger-scale perturbations for stratiform events and smaller-scale ones for convection, a sign that the models are responding to genuine dynamical differences between the two regimes.

The eigenvalue growth itself followed power laws in lead time, a behaviour consistent with the theoretical predictability limits of atmospheric flows whose energy spectra follow the classical minus five-thirds power law at nowcasting scales. Errors seeded at small scales cascade toward larger ones, and the perturbation modes gradually develop power at progressively larger scales as the forecast advances. This hierarchy, in which the largest-amplitude perturbations affect the largest scales, mirrors the way predictability is lost in the real atmosphere, and it is encouraging that both a physics-inspired statistical scheme and a purely data-driven generative model reproduce it.

The critical test came when the researchers asked whether the ensembles could localise the error in space, not just quantify it. They used two complementary metrics: the cosine of the angle between the actual forecast error and its projection onto the ensemble perturbations, a geometric measure related to the PECA score, and the probabilistic Fraction Skill Score, which assesses spatial accuracy while tolerating displacement errors. The clever twist was the construction of surrogate ensembles. MAAFT surrogates, inspired by the Iterated Amplitude-Adjusted Fourier Transform, preserve the statistical fingerprints of the original ensembles, namely the distribution of pixel values and the power spectra of the perturbations, but scramble the complex Fourier phases that encode spatial structure. If the real ensembles contain genuine spatial information, they should outscore these statistical ghosts.

They did not. The scores of the MAAFT surrogates were essentially indistinguishable from those of the actual STEPS and LDCast ensembles for both metrics, at all lead times and for both event types. In Fourier language, the complex phases of the ensemble perturbations, which carry all the spatial localisation information, are effectively random. The ensembles know how big the error should be at each scale, and they even reproduce the right distribution of rainfall intensities, but they have no dynamical insight into where the error will actually materialise. The skill they display is statistical, not dynamical. A similar picture emerged from the Fraction Skill Score analysis, where both models’ scores saturated with lead time toward values achievable by purely random predictions constrained to the same value distributions.

The authors are careful about what this means for operations. The finding does not invalidate generative nowcasting; LDCast’s spread still provides a useful, nearly unbiased estimate of error magnitude across most scales, even without regional retraining, and both models demonstrably adapt their uncertainty to the weather regime. But forecasters hoping that an ensemble of AI rainfields will point toward the specific locations where the forecast is going wrong will be disappointed. The study also notes a practical asymmetry: when STEPS runs out of radar data at the edge of its domain, it produces explicit missing values that honestly flag where information is lacking, whereas a generative model like LDCast simply invents plausible-looking values with no such transparency. Future work, the authors suggest, should test a version of LDCast fine-tuned for Belgium, evaluate other generative architectures such as DGMR, which produces less diverse members, and explore blending nowcasting models with numerical weather prediction, an approach already available in pysteps. For now, the message is clear: AI ensembles have learned the statistics of rainfall uncertainty, but the dynamics of where that uncertainty strikes remain, for the moment, beyond their grasp.

Subject of Research: Evaluation of ensemble spread-error relationships and spatial error representation in precipitation nowcasting using STEPS and the generative AI model LDCast over Belgium

Article Title: Spread/error relationship and spatial error representation in precipitation nowcasting: comparison of STEPS and generative AI

Article References: Bonte, M., De Cruz, L., Debal, F., & Vannitsem, S. (2026). Spread/error relationship and spatial error representation in precipitation nowcasting: comparison of STEPS and generative AI. Weather and Climate Dynamics, 7(3), 1837-1852. https://doi.org/10.5194/wcd-7-1837-2026

Image Credits: AI Generated

DOI: 10.5194/wcd-7-1837-2026

Keywords: precipitation nowcasting, LDCast, STEPS, generative AI, ensemble forecasting, spread-error relationship, latent diffusion model, radar, uncertainty quantification, Fraction Skill Score, convective precipitation, Weather and Climate Dynamics

The work was led by Sharan Majumdar, a professor of atmospheric sciences at the University of Miami’s Rosenstiel School of Marine, Atmospheric, and Earth Science, in collaboration with scientists from ECMWF, the NSF National Center for Atmospheric Research, and the University of Bonn. Among the collaborators was Quinton Lawton, a former doctoral student of Majumdar’s who is now an assistant professor in the Department of Earth, Atmosphere and Environment at Northern Illinois University. The team focused their analysis on the Atlantic basin from 2020 through 2024, a five-year window that captures both steady improvements in conventional forecasting and the dramatic arrival of AI-based alternatives.

At the heart of the study is a comparison between two ECMWF systems. The first is the Integrated Forecasting System, or IFS, the center’s traditional numerical weather prediction model, which simulates the atmosphere by solving the physical equations that govern fluid motion, thermodynamics, and moisture on a three-dimensional grid. The second is the Artificial Intelligence Forecasting System, or AIFS, a machine-learning model trained on decades of historical atmospheric data to learn the patterns of atmospheric evolution directly. Rather than stepping forward through time with explicit physics calculations, the AIFS produces forecasts in a fundamentally different way, and ECMWF made it operational in 2025, marking a significant expansion of machine-learning-based forecasting alongside its physics-based numerical models.

The specific targets of the analysis were African easterly waves, the westward-moving atmospheric disturbances that form over sub-Saharan Africa during the boreal summer. These waves are the seedlings of many Atlantic hurricanes. As they roll off the West African coast and cross the tropical Atlantic, only some of them manage to organize into tropical cyclones, and identifying which ones will succeed is a central challenge of hurricane forecasting. The timing and location of this transition, known as tropical cyclogenesis, can determine how much warning coastal populations receive, which makes any improvement in lead time potentially life-saving.

The researchers first established a baseline by examining how the conventional IFS performed across the study period, and the results show genuine progress. In 2023, the grid spacing of the IFS ensemble was reduced from 18 kilometers to 9 kilometers, effectively doubling the model’s resolution and allowing it to represent smaller-scale tropical features with greater fidelity. Over the same years, the average probabilities of tropical cyclone formation produced by the conventional system generally increased, indicating growing confidence and skill in detecting developing systems. These improvements reflect the steady, incremental gains that operational forecasting centers have delivered through better observations, more sophisticated model physics, and finer computational grids.

But the more striking findings emerged when the team turned to the AI system. Among the 18 tropical cyclones that developed in the Atlantic in 2024, the AIFS ensemble frequently produced higher probabilities of development than the IFS at lead times of roughly 84 to 120 hours, which corresponds to about three and a half to five days before formation. This advantage was particularly pronounced for stronger tropical waves, the very systems that forecasters most want to flag early. In other words, when a disturbance showed signs of being a serious candidate for development, the AI model often recognized that potential days sooner than the traditional ensemble did.

The picture reversed at shorter lead times. At 36 to 48 hours before formation, the AIFS ensemble generally produced lower probabilities of development than the IFS, especially for weaker systems. This contrast highlights one of the fundamental challenges of forecasting tropical cyclogenesis: the signal that a disturbance will become a tropical cyclone can change substantially as the event approaches. A wave that looked promising five days out may struggle to consolidate, while a seemingly innocuous disturbance can rapidly organize. The study also found that conventional IFS probabilities often increased sharply when forecasts moved from three days to two days before a storm was officially named, and that there was considerable variation from storm to storm, depending in part on the wind-speed threshold used to define development. Cyclogenesis, it turns out, is not a single crisp event but a process whose detectability depends on both the system and the definition applied to it.

The AI system also showed advantages in forecasting where developing tropical systems would be located. The average position error of the AIFS ensemble mean was often smaller than the errors from the AIFS single forecast, the deterministic IFS, and the IFS ensemble mean. Position accuracy matters enormously in practice, because a disturbance’s location relative to warm ocean waters, wind shear, and dry air strongly influences its chances of survival, and because forecasters use position guidance to decide where to focus monitoring resources, aircraft reconnaissance, and public communications. An ensemble mean that consistently pins down the location of a developing system gives forecasters a more reliable anchor for their assessments.

Importantly, the findings do not suggest that AI should replace traditional numerical weather prediction. Majumdar emphasized that AI forecasts can complement the established IFS, providing forecasters with another source of information when assessing the potential for tropical cyclone formation. This framing reflects a growing consensus in the operational meteorology community: machine-learning models trained on historical reanalyses can capture large-scale patterns with remarkable efficiency and at a fraction of the computational cost of physics-based simulation, but the two approaches have different failure modes, and agreement or disagreement between them is itself informative. A forecaster seeing high development probabilities from both systems at a four-day lead time has a very different situation than one seeing a divergence between the two.

The study was supported by a National Science Foundation grant, AGS-2438140, and by the European Centre for Medium-Range Weather Forecasts, and it arrives at a moment of rapid institutional change. ECMWF’s decision to make the AIFS operational in 2025 signals that machine-learning forecasting has moved from research curiosity to production tool at one of the world’s leading centers. For Atlantic hurricane forecasting, Majumdar pointed out, the potential significance is substantial. Better identification of developing systems several days in advance, he said, could give forecasters more time to monitor disturbances, assess possible tracks and communicate emerging risks. The study results point toward a future in which traditional physics-based forecasting and AI-generated guidance work side by side or are combined, potentially giving forecasters a more complete picture of which Atlantic tropical waves are most likely to become the next named storm. As the 2020s progress and both classes of models continue to improve, the practical question is shifting from whether AI belongs in the hurricane forecaster’s toolkit to how best to blend its strengths with the physics-based systems that have anchored tropical cyclone prediction for decades.

Subject of Research: Artificial intelligence and numerical model forecasting of Atlantic tropical cyclogenesis from African easterly waves

Article Title: AI’s eye on the tropics

Article References: AI’s eye on the tropics. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: tropical cyclones, hurricane forecasting, artificial intelligence, ECMWF, African easterly waves, tropical cyclogenesis, ensemble forecasting, machine learning, Atlantic basin, Weather and Forecasting, numerical weather prediction, University of Miami

SURF was developed by Mathew J. Owens and Luke A. Barnard of the University of Reading and described in the journal Solar Physics. The framework packages together two modelling options. The first is HUXt, a well-established reduced-physics model that treats the solar wind as a one-dimensional advection problem and has already found use in both forecasting and a diverse range of scientific applications, from planetary studies to comet-tail analysis. The second, and the centrepiece of the new work, is a newly developed one-dimensional compressible hydrodynamic solver called hydro, which adds physically consistent compression effects while retaining the computational efficiency needed for ensemble forecasting, uncertainty quantification and large parametric studies.

The distinction matters because the solar wind is a compressible fluid. Fast streams emitted from coronal holes eventually catch up with slower wind ahead of them, piling plasma into compressed stream interaction regions bounded by shock-like fronts. HUXt, which lacks physics-based compressibility, agrees with full three-dimensional magnetohydrodynamic models to within about five percent for the same boundary conditions, but its largest deviations appear precisely at those compression fronts. SURF-hydro addresses this weakness by solving the one-dimensional Euler equations of mass, momentum and energy conservation in spherical geometry, using finite-volume methods with Riemann solvers and second-order spatial reconstruction.

The numerical machinery is sophisticated but conceptually standard in computational fluid dynamics. The model discretises the radial domain into spherical shell cells and computes fluxes across cell interfaces using a Harten-Lax-van Leer-Contact solver, which handles the discontinuities that arise at shocks. Two reconstruction schemes are available: a first-order piecewise constant method for maximum speed, and a default second-order piecewise linear method with a monotonized central limiter that keeps sharp features like shocks intact without introducing spurious oscillations. Because the flow is radial in spherical geometry, the model includes a geometric source term in the momentum equation, reflecting the fact that the surface area of a sphere grows as the square of distance from the Sun.

Validation against an analytical benchmark is impressive. The team compared SURF-hydro with an exact solution for steady-state, pressure-driven expansion of a uniform spherical wind, treating the solar wind as isentropic flow through a nozzle whose cross-sectional area grows with distance squared. The second-order solution reproduced the analytical solar wind speed with an error of just 0.04 percent, with density and temperature errors below one percent, and conserved mass to within one percent across the whole domain. Speed matters too: a five-day simulation for a single longitude takes about 0.1 seconds on a standard desktop processor, roughly ten thousand times cheaper than a full three-dimensional magnetohydrodynamic run.

Perhaps the most consequential contribution is a new way of setting the model’s inner boundary conditions. Solar wind models typically start at 0.1 astronomical units, about a fifth of Mercury’s orbital distance, where the flow is already super-magnetosonic. Coronal models supply speed and magnetic field at that boundary, but density and temperature must be inferred from the speed, usually by assuming some form of equilibrium such as constant mass, momentum or kinetic energy flux. Owens and Barnard instead mined thirty years of near-Earth OMNI observations, removed all periods contaminated by coronal mass ejections, and derived empirical relations between solar wind speed, density and temperature at 1 astronomical unit. They then back-mapped those relations to 0.1 astronomical units using the analytical nozzle solution, producing a non-equilibrium look-up table that can be interpolated for any inner-boundary speed.

The payoff shows up in hindcast tests. For a representative 27-day interval of recurrent solar wind in 2019, a SURF-hydro hindcast driven by back-mapped in situ observations reproduced the observed base-level proton density of around five particles per cubic centimetre, along with the sharp density spikes of tens of particles per cubic centimetre at stream interaction regions, and temperatures ranging from about 50,000 kelvin in slow wind to 500,000 kelvin in compressed regions. By comparison, archived operational WSA-Enlil forecasts for the same period showed almost no density variation and temperatures systematically an order of magnitude too low. Crucially, the authors show this is not a flaw in Enlil’s physics but in its boundary conditions: when the same WSA coronal maps drove SURF-hydro with the new non-equilibrium relations, the variability in speed, density and temperature all improved markedly.

Extending the comparison across four years of observations reinforced the point. WSA-Enlil systematically under-dispersed solar wind speeds and produced far too little density variability, while its temperatures remained far too low even accounting for the reduced speed range, implicating the equilibrium assumption at the inner boundary. WSA-SURF-hydro, using the empirical relations, matched the observed ranges and trends much more closely, with the main discrepancy being somewhat elevated densities and temperatures at intermediate speeds, likely because the WSA coronal model produces too many fast streams. The authors suggest that operational systems could be significantly improved simply by adopting similar empirical density and temperature relations, an approach transferable to other solar wind models beyond SURF.

The framework also shines a light on an under-explored source of forecast uncertainty: the assumed properties of coronal mass ejections themselves. Operational systems insert CME perturbations at 0.1 astronomical units that are over-dense, typically four times the ambient density, partly to compensate for the neglected internal magnetic pressure of the cone-model representation. Yet observations at 1 astronomical unit show that interplanetary coronal mass ejections are actually cooler and more tenuous than the surrounding wind, partly from adiabatic expansion in transit and partly because significant expansion and cooling has already occurred close to the Sun. A super-posed epoch analysis of 45 fast magnetic clouds confirmed this picture, with the ejecta body characterised by declining speed and lower density and temperature than the ambient solar wind.

Because SURF-hydro can sample parameter space rapidly, the team ran sensitivity tests varying the initial density and temperature of a model CME launched into a structured ambient wind. Even without magnetic forces, the model reproduced the key observed features of CME evolution at 1 astronomical unit, including a hot, dense sheath ahead of the ejecta, an expanding body cooler and less dense than its surroundings, and durations of roughly 24 hours consistent with observations. The sensitivity results were striking: for one particular structured solar wind, varying the CME’s initial density and temperature at 0.1 astronomical units changed the transit time and arrival speed at Earth by 15 to 20 percent. Hotter, denser CMEs arrived sooner, faster and with stronger shocks. Since these parameters are observationally unconstrained and currently ignored in ensemble forecasting, the authors argue they merit systematic perturbation in future operational ensembles. With the SURF code freely available on GitHub and installable from PyPI and conda-forge, the framework offers researchers and forecasters alike an efficient bridge between idealised models and full three-dimensional simulations, and a practical tool for interrogating the assumptions that quietly shape every space-weather forecast.

Subject of Research: Compressible hydrodynamic modelling of the solar wind and coronal mass ejection propagation for space-weather research and forecasting

Article Title: Space-Weather Utilities for Research and Forecasting (SURF): A Tool for Investigating Hydrodynamic Aspects of Solar Wind and Coronal Mass Ejection Expansion and Evolution

Article References: Owens, M. J., & Barnard, L. A. (2026). Space-Weather Utilities for Research and Forecasting (SURF): A Tool for Investigating Hydrodynamic Aspects of Solar Wind and Coronal Mass Ejection Expansion and Evolution. Solar Physics, 301(9), Article 147. https://doi.org/10.1007/s11207-026-02738-7

Image Credits: AI Generated

DOI: 10.1007/s11207-026-02738-7

Keywords: space weather, solar wind, coronal mass ejections, SURF, HUXt, hydrodynamic modelling, OMNI observations, WSA-Enlil, forecasting, heliospheric physics, boundary conditions, ensemble forecasting

For forecasters, events like this are among the hardest problems in numerical weather prediction. The northwestern Himalaya presents a brutal combination of challenges: elevations that climb to roughly six kilometers, sparse surface observations, unreliable satellite rainfall estimates, and a limited radar network. Global and regional models struggle to represent how moisture converges, how convection organizes, and how the terrain itself modulates precipitation when the grid cells of a model are coarse compared with the landscape they are trying to describe. A new study published in Discover Geoscience by Vijay Vishwakarma, Chandra Shekhar Satapathy, and Sandeep Pattnaik of IIT Bhubaneswar, together with Rajendra Jenamani of the India Meteorological Department and Mihir Kumar Das, argues that a crucial piece of the puzzle lies not in the sky but underfoot: the initial state of the land surface, and in particular the soil moisture with which a model begins its forecast.

The research team used the Weather Research and Forecasting model, version 4.2.1, in a two-nest configuration with grid spacings of 9 and 3 kilometers and 35 vertical levels reaching to 50 hectopascals. Initial and lateral boundary conditions came from the ERA5 global reanalysis of the European Centre for Medium-Range Weather Forecasts at 0.25 degree resolution. The central innovation was the use of the High-Resolution Land Data Assimilation System, or HRLDAS, version 3.7.1, which was integrated continuously from January 2008 through August 2023 using forcing data from the Global Land Data Assimilation System. This long spin-up allowed realistic, internally consistent fields of soil moisture, soil temperature, and canopy water content to evolve over the Indian region, which were then used to replace the default soil moisture in the model’s initial conditions. In parallel, the team compared two land use and land cover datasets: the long-standing global USGS classification and a far more current Indian dataset derived from the Indian Space Research Organisation’s Resourcesat-1 Advanced Wide Field Sensor, which classifies Indian land cover at 56 meter native resolution and captures recent changes in vegetation, agriculture, and urbanization.

The experimental design was deliberately exhaustive. Six different parameterization schemes were selected for each of three families of model physics: cloud microphysics, cumulus convection, and the planetary boundary layer. Combined with the two land cover datasets and the two land initialization approaches, this yielded 72 numerical simulations spanning lead times of up to four days. These were organized into eight control ensembles, which used conventional land states, and eight HRLDAS ensembles, which used the assimilated land states. Forecast skill was evaluated against daily rainfall from the India Meteorological Department at 0.25 degree resolution, station-level observations from 15 gauges in Himachal Pradesh situated between roughly 850 and 3,000 meters elevation, and satellite retrievals from the Global Precipitation Measurement mission’s IMERG product. The primary verification metric was the absolute percentage error between simulated and observed 24-hour accumulated rainfall.

The results were striking. On the peak day of the event, the ensembles built on ISRO land cover data and HRLDAS soil moisture achieved an absolute percentage error of about 32 percent across the Himachal Pradesh stations, while the control ensembles exceeded 40 percent. By day 3 of the forecast, as the event decayed, the HRLDAS ensembles maintained errors near 20 percent, whereas the control ensembles drifted above 50 percent, unrealistically amplifying their errors as the simulation progressed. The improvement was most pronounced at low-elevation stations, where the model’s 3-kilometer grid can better resolve the sub-grid processes that generate rain. District-scale analysis using Taylor diagrams reinforced the picture: HRLDAS ensembles showed stronger pattern correlations, above 0.3 and reaching roughly 0.6 for the best configurations, and lower normalized standard deviations than their control counterparts, which frequently doubled the observed spatial variability.

The study also quantified which physical processes matter most over this terrain. Ensembles grouped by microphysics scheme produced the most realistic rainfall distributions, followed by those grouped by cumulus parameterization, with planetary boundary layer ensembles performing weakest. This ordering suggests that the representation of cloud and precipitation microphysics, the processes by which vapor condenses into droplets and ice and grows into rain and snow, exerts dominant control over simulated rainfall in the Himalaya, more so than the treatment of turbulent mixing in the boundary layer. Categorical verification using the equitable threat score and probability of detection showed the HRLDAS ensembles peaking on day 2, with threat scores near 0.2 and detection probabilities near 0.5 for heavy rainfall thresholds, while cumulative distribution functions and relative operating characteristic curves confirmed that the ISRO-based HRLDAS ensembles aligned most closely with observations.

Perhaps the most physically illuminating part of the analysis examined the vertical structure of the simulated storms using contoured frequency by altitude diagrams of radar reflectivity and hydrometeors. During the peak 24 to 72 hours of the event, the HRLDAS ensembles displayed enhanced reflectivity at mid-to-upper altitudes, indicating more robust convective activity, along with a more realistic vertical partitioning of hydrometeors: more liquid water in the lower troposphere and more ice-phase particles aloft. This vertical arrangement is exactly what is expected in organized monsoon convection over mountains, where warm-rain processes low in the cloud and mixed-phase processes higher up jointly determine how much rain reaches the surface. The control ensembles, by contrast, misplaced hydrometeor mass in the vertical, overproducing both liquid and solid species in the mid-troposphere. The authors interpret this as evidence that better land initialization strengthens the land-atmosphere coupling that feeds moisture and instability into the convective systems responsible for the disaster.

The study is candid about remaining failures. Every ensemble, regardless of configuration, underestimated rainfall over the high-altitude districts of Himachal Pradesh, including Chamba, Kullu, and Lahul and Spiti, which sit above 3,000 meters and registered rainfall vulnerability index values above 0.75 in observations. The rainfall vulnerability index, a normalized measure of district-scale rainfall intensity, agreed well with satellite observations across most of Uttarakhand, where terrain is comparatively smoother and lower, but the models could not replicate the observed vulnerability of the highest Himachal districts. The authors attribute this to sub-grid-scale terrain-induced processes that remain unresolved even at 3-kilometer resolution, compounded by coarse forcing data and the scarcity of high-altitude observations. They argue that dense networks of meteorological observatories in these locations, along with convection-permitting finer-resolution configurations, radar and satellite data assimilation, and coupled atmosphere-land assimilation frameworks, are needed to close the gap.

The broader significance of the work extends well beyond a single disaster. Heavy rainfall events are projected to intensify under global warming, following the Clausius-Clapeyron relationship by which a warmer atmosphere holds more water vapor, and the Himalayan states, with their dense biodiversity, critical infrastructure, and thriving tourism economy, are acutely exposed. The finding that assimilated, long-spin-up soil moisture combined with an up-to-date national land cover dataset can cut forecast errors by nearly a third on the peak day of a catastrophic event carries direct operational weight. The authors suggest that HRLDAS-based land initialization is ready for implementation within operational numerical weather prediction systems, where it could strengthen impact-based early warning, give disaster response agencies more lead time, and ultimately save lives in one of the world’s most rainfall-vulnerable mountain regions. The message is simple and consequential: to predict what falls from Himalayan skies, forecasters must first know what lies beneath them.

Subject of Research: Improving ensemble-based heavy rainfall forecasting over the Indian northwestern Himalaya using high-resolution land data assimilation and updated land use datasets

Article Title: Improvement in ensemble-based localized rainfall forecast skills over the Indian northwestern Himalaya states using initial states from a high-resolution land data assimilation system

Article References: Vishwakarma, V., Satapathy, C. S., Pattnaik, S., Jenamani, R., & Das, M. K. (2026). Improvement in ensemble-based localized rainfall forecast skills over the Indian northwestern Himalaya states using initial states from a high-resolution land data assimilation system. Discover Geoscience, 4(1), Article 368. https://doi.org/10.1007/s44288-026-00746-5

Image Credits: AI Generated

DOI: 10.1007/s44288-026-00746-5

Keywords: heavy rainfall events, WRF model, HRLDAS, soil moisture, land use land cover, Himalaya, monsoon, ensemble forecasting, numerical weather prediction, Himachal Pradesh, Uttarakhand, disaster preparedness

The work was carried out by C.W.E. de Korte, M. Verlaan, A. W. Heemink and B. Backeberg, affiliated with Delft University of Technology and the research institute Deltares. Their starting point was a persistent problem: third-generation wave models such as SWAN, the Simulating WAves Nearshore model used in the study, are highly reliable in the open ocean but continue to struggle in coastal shelf seas. Wave-current interactions, uncertain wind forcing, shallow-water effects and imperfect parametrizations of the physical source terms all introduce errors that are difficult to eliminate by calibration alone. Data assimilation offers a different route, blending real observations with the model’s own physics to nudge the simulated ocean state closer to reality without rewriting the underlying equations.

What sets this study apart from most earlier wave data assimilation efforts is the choice of state variable. Conventional operational schemes, such as Optimal Interpolation and three-dimensional variational methods, typically apply corrections only to significant wave height and then scale those corrections back onto the full wave spectrum using simplifying assumptions about how wave energy is distributed across frequencies and directions. The Dutch team instead placed the complete directional wave energy spectrum in the model state. Each ensemble member carried the spectrum across 32 frequency bands and 36 directional bins at every point of a 567-cell grid, producing a state vector of more than 650,000 elements per member. Because the ensemble evolves under the full SWAN physics, the corrections the filter produces are automatically physically consistent with the model, and integral parameters such as mean wave period adjust themselves without any ad hoc scaling.

The DEnKF itself is a deterministic variant of the classic Ensemble Kalman Filter. Rather than perturbing observations with random noise to propagate uncertainty, it updates the ensemble mean and the ensemble anomalies separately, avoiding the sampling errors that stochastic perturbations introduce. This is particularly valuable for small ensembles, and the team settled on 64 members after previous synthetic twin experiments showed the error statistics fully converged at that size. Uncertainty was injected into the system through the wind forcing, treated as the control variable, using a first-order autoregressive noise model with a spatial Gaussian correlation structure. Parameters were derived from the difference between HARMONIE wind analyses and forecasts, giving a standard deviation of two metres per second, a decorrelation timescale of fifteen hours and a spatial decorrelation length of 500 kilometres.

The observations came from seven nadir satellite altimeters: CFOSAT, Haiyang-2B, Cryosat-2, Jason-3, the two Sentinel-3 satellites, and Saral/AltiKa, all retrieved from the Copernicus Marine Environment Monitoring Service. Over the three-month window the satellites contributed 713 tracks over the North Sea, an average of about eight passes per day, with a mean interval of roughly three hours between passes but gaps stretching to nearly fifteen hours. Tracks were sub-sampled every 120 kilometres to avoid overloading individual grid cells, and a coastal mask excluded measurements within 50 kilometres of shore, where altimeter retrievals are known to be unreliable. Observation errors were assumed to be uncorrelated with a standard deviation of 0.2 metres. Hamill localisation experiments comparing the standard EnKF with the DEnKF across localisation radii of 100 to 500 kilometres showed the DEnKF with a 200-kilometre radius performed best, and that configuration became the final set-up.

Validation against the North Sea’s dense network of independent wave buoys delivered strikingly consistent results. Significant wave height errors dropped by a mean of 20.5 percent, from 0.39 metres in the free-running coarse model to 0.31 metres, a performance essentially matching the much finer SWAN-DCSM benchmark model run at roughly 3.6-kilometre resolution. Mean wave period improved at 21 of 23 stations with a ten percent reduction in root mean square error, while the peak period improved at 13 of 16 stations by about five percent. Wind speed showed modest improvements at some stations, though the researchers caution that the station anemometers, corrected to ten-metre equivalent heights assuming a neutral wind profile, carry their own uncertainties over the frequently non-neutral marine boundary layer. Not every parameter benefited: swell wave height degraded slightly on average, and the low-frequency inverse moment period and mean wave direction, each measured at only a handful of stations, also worsened marginally.

Spectral analysis explained the pattern. In unimodal sea states dominated by wind-driven waves, the assimilation corrected the entire wave spectrum in a way that closely matched buoy observations, as demonstrated during a storm peak on 20 January 2022 at the offshore station A121, where the analysis tracks the measured spectra hour by hour. But in mixed sea states where wind-sea and swell are clearly separated, the ensemble spread remained concentrated in the mid and high frequencies, because wind perturbations barely touch an independently propagating swell field and the altimeters measure only total significant wave height. Detailed examination of the largest swell errors revealed two distinct mechanisms: during short-fetch, rapidly rotating local wind conditions, the wind-based error covariances failed to represent the spatial scales at which swell actually varied between neighbouring stations, while a second error type, premature swell arrival at coastal stations, proved to be a systematic bias of the coarse-resolution model rather than a failure of the assimilation itself.

One of the most practically important findings concerns timing. The researchers binned all validation samples by the number of hours elapsed since the last satellite pass and found that prediction errors rose steadily with the length of the gap, particularly for stations in the open central North Sea. Coastal stations, whose errors are dominated by shallow-water processes rather than wind-driven corrections, were less sensitive to the satellite schedule. Because the orbital geometry of the contributing satellites fixes the timing of the gaps, these gaps recur with the tidal cycle, a phase-locking effect the authors flag as deserving further study. The message for forecasters is clear: the temporal density of observations matters, and merging additional data sources could deliver substantial gains.

The authors are candid about the limitations. The coarse 0.5-degree grid, chosen so that a 64-member ensemble of full spectra could be run at all, degrades accuracy near the coast, where resolution, missing triad interactions and a simplified setup all take their toll. Running a high-resolution model within the ensemble framework would require major advances in computing power and memory. Still, the path forward is mapped out: refining the wind noise model, assimilating additional integral wave parameters or even full spectra, incorporating continuous buoy measurements, and adding satellite SAR observations from missions such as Sentinel-1, SWOT and Sentinel-6. Beyond operational forecasting, the team points to wave reanalyses for risk and climate studies, and to the growing demand for high-quality training data for machine-learning wave models. For a shelf sea as busy and economically vital as the North Sea, better wave information from the satellites already overhead is a prize worth the computation.

Subject of Research: Ensemble-based assimilation of satellite altimeter wave measurements in a regional North Sea wave model

Article Title: Exploring the added value of assimilating satellite altimeter measurements in a North Sea wave model using the Deterministic Ensemble Kalman Filter

Article References: de Korte, C., Verlaan, M., Heemink, A. W., & Backeberg, B. (2026). Exploring the added value of assimilating satellite altimeter measurements in a North Sea wave model using the Deterministic Ensemble Kalman Filter. Ocean Dynamics, 76(10), Article 102. https://doi.org/10.1007/s10236-026-01858-9

Image Credits: AI Generated

DOI: 10.1007/s10236-026-01858-9

Keywords: wave data assimilation, Deterministic Ensemble Kalman Filter, satellite altimeter, SWAN wave model, North Sea, significant wave height, wave spectrum, ensemble forecasting, swell, coastal shelf seas, Ocean Dynamics, CMEMS

The research, led by Mohammadhossein Homaei of the University of Extremadura together with colleagues at Bowling Green State University, introduces CAUCCES, a digital twin framework that couples an ensemble of four forecasting models with a novel uncertainty measure called the Explainable Confidence Index, or ECI. The team validated the system across twelve Spanish municipalities over eighteen months, and the headline result is striking: ECI-driven scheduling cut SLA violations from 9.3 percent to 1.5 percent, while adding only 3.1 percent to operational costs. In an industry where a single service interruption can mean regulatory fines and eroded public trust, that trade-off is remarkable.

The core problem the researchers set out to solve is a stubborn gap between forecasting and scheduling. Existing digital twin platforms for water networks typically feed deterministic point forecasts into maintenance schedulers, treating the predicted demand as if it were certain. When actual demand exceeds the prediction while a pipeline is under maintenance, hydraulic constraints are violated and customers experience service interruptions. Rigorous Bayesian uncertainty methods could quantify that risk, but they come with a punishing computational price: the study measured a Bayesian LSTM baseline at 340 milliseconds per inference, far too slow for real-time scheduling on the standard hardware most utilities can afford.

CAUCCES sidesteps that bottleneck with an adaptive ensemble of deliberately diverse models: a dual-branch LSTM network that ingests meteorological data alongside consumption history, a Prophet model configured with weather regressors and multiplicative seasonality, and two gradient boosting learners, LightGBM and XGBoost. The ensemble achieved a mean absolute percentage error of 14.12 percent, outperforming DeepAR at 16.50 percent and the Temporal Fusion Transformer at 16.92 percent. The authors attribute this advantage to architectural diversity rather than raw model depth. Each member captures a different pattern: the LSTM handles short-term recency, Prophet captures annual seasonality through Fourier decomposition, and the boosting models capture non-linear temperature thresholds. For rural utilities with limited historical data, that diversity proves more reliable than scale.

The real innovation, however, is the ECI itself. Rather than requiring hundreds of Monte Carlo forward passes, the index is computed in closed form from two signals already present in any ensemble: the spread of predictions across models, which serves as a proxy for irreducible variability, and the normalized Shannon entropy of the ensemble weights, which captures disagreement among the models about the underlying demand pattern. The two components are combined multiplicatively, a choice validated by cross-validation, because periods where both spread and disagreement are simultaneously elevated accounted for 78 percent of observed scheduling failures. The resulting score is normalized by a rolling 30-day variance percentile, allowing the metric to adapt to seasonal shifts between volatile summers and stable winters.

Crucially, the ECI is not merely a diagnostic. It plugs directly into a constraint programming scheduler as a risk penalty: tasks scheduled during low-confidence windows incur exponentially rising costs, so the solver defers non-critical work such as routine inspections or meter replacements until confidence recovers. True emergency repairs always follow strict priority rules and are never deferred. In operational terms, an ECI above 0.8 means schedule everything; between 0.6 and 0.8, defer non-critical tasks; below 0.6, postpone all non-emergency maintenance. The researchers proved mathematically that under an active risk budget, the optimizer is guaranteed to postpone tasks during windows where confidence falls below a critical threshold, and field data confirmed the prediction: 68 percent of low-importance tasks were deferred when ECI dropped below 0.6.

The team also subjected the index to a battery of validation experiments. Against a Bayesian LSTM trained with variational inference, ECI-derived prediction intervals showed a Pearson correlation of 0.76 with Bayesian credible interval widths, while running 28 times faster, at 12 milliseconds versus 340. On 500 held-out samples, empirical coverage matched target coverage within 1.4 percent on average. An ablation study confirmed that both components matter: variance alone yielded a 4.2 percent failure rate, entropy alone 6.1 percent, and the full ECI 1.5 percent, approaching an oracle bound of 0.8 percent that assumes perfect foresight of demand surges.

The broader operational gains are equally notable. Across the deployment, the platform reduced task completion time by 14 percent, emergency response time by 25 percent, customer service disruption by 31 percent, and carbon dioxide emissions by 17 percent, with fuel consumption down 16 percent. Field telemetry from OBD-II sensors and GPS trackers on the maintenance fleet confirmed the calculated environmental improvements within 1.2 percent. The multi-objective optimizer balances completion time, fuel, emissions, and customer impact, with weights chosen from the knee point of a Pareto front mapped using the epsilon-constraint method, ensuring the chosen operating point is non-dominated.

The study is candid about its limits. Multi-regional testing across Andalusia, Catalonia, and the Basque Country revealed performance degradation of 12 to 26 percent outside the development region of Extremadura, a consequence of climate-specific feature engineering and ensemble weights. The authors prescribe a recalibration protocol, at least 90 days of local data, retrained ensemble weights, and locally recalibrated variance normalization, which reduces degradation to 5 to 8 percent. The constraint solver also reaches practical limits around 300 daily tasks, making the framework best suited to rural and small-urban utilities rather than megacities. A three-day sensor outage during the study pushed forecast error from 14 to 42 percent, underscoring the system’s dependence on continuous telemetry.

Even with those caveats, the implications reach well beyond Spanish water networks. Roughly 60 percent of European water infrastructure serves populations under 50,000, utilities for which commercial enterprise systems are economically out of reach. CAUCCES runs its inference and scheduling on commodity hardware, with no GPU required at deployment, and the authors have released code and anonymized datasets for community testing. The deeper lesson may be methodological: uncertainty is not a footnote to forecasting but a first-class input to operations. By converting ensemble disagreement into a number a scheduler can act on, the study offers resource-constrained utilities a practical bridge between what their models know and what their crews should do, and it suggests that knowing when not to act may be the most valuable prediction of all.

Subject of Research: Uncertainty-aware maintenance scheduling in water distribution networks using ensemble neural forecasting and explainable confidence indexing

Article Title: Uncertainty-aware maintenance scheduling in water distribution networks via ensemble neural forecasting and explainable confidence indexing

Article References: Homaei, M., Mogollon-Gutierrez, O., Rezaee, M. M., Caro, A., & Avila, M. (2026). Uncertainty-aware maintenance scheduling in water distribution networks via ensemble neural forecasting and explainable confidence indexing. Neural Computing and Applications, 38(17), Article 733. https://doi.org/10.1007/s00521-026-12351-1

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12351-1

Keywords: water distribution networks, digital twin, ensemble forecasting, uncertainty quantification, maintenance scheduling, machine learning, LSTM, explainable AI, smart water management, SLA violations, constraint programming, climate variability