Recent developments, including the integration of AI and machine learning applications, the emergence of kilometre-scale simulations, and the widespread availability of open-source downscaling products, add new dimensions to this challenge. These advances raise important questions about the ‘added value’ of downscaling, especially in light of the cascade of uncertainty and the need for robust evaluation frameworks.
In our session, we aim to bring together scientists from the various geoscientific disciplines interrelated through downscaling: atmospheric modeling, climate change impact modeling, machine learning and verification research. We also invite philosophers of climate science to stimulate our discussion about the novel challenges that arise from evaluating complex models and modelling chains in the face of the increasingly heterogeneous needs of the growing user communities.
Contributions to this session may address, but are not limited to:
- newly available downscaling products,
- applications relying on downscaled data and impact assessments,
- downscaling method development and machine learning,
- bias correction and statistical postprocessing,
- challenges in the data management of kilometer-scale simulations,
- verification, uncertainty quantification and the added value of downscaling,
- downscaling approaches in light of computational epistemology.
Climate change is amplifying extreme events, posing escalating risks to biodiversity, human health, and food security. Global climate models (GCMs) are essential for projecting future climate, yet their coarse resolution and high computational costs constrain their ability to represent extremes. Here, we introduce FuXi-CMIPAlign, a generative deep learning framework for downscaling Coupled Model Intercomparison Project (CMIP) outputs. The model integrates Flow Matching for generative modeling with domain adaptation via Maximum Mean Discrepancy loss to align feature distributions between training data (ERA5 reanalysis) and inference data (European Consortium-Earth), thereby mitigating input discrepancies and improving accuracy, stability, and generalization across emission scenarios. FuXi-CMIPAlign performs spatial, temporal, and multivariate downscaling, enabling more realistic simulation of compound extremes such as tropical cyclones (TCs). Applied to the historical period (2005–2014), it reduces global 99th-percentile mean absolute errors by 26%, 42%, and 33% for high temperature, extreme precipitation, and strong wind, respectively, and reproduces TC activity better aligned with ERA5. Under future scenarios (2015–2100), FuXi-CMIPAlign projects pronounced increases in land area affected by high temperature and frequency of extreme precipitation under high-emission scenarios, along with up to 60% rise in TC intensity and frequency over the Northwest and Northeast Pacific. In contrast, strong wind events over land shows a counterintuitive weakening trend. These results demonstrate that FuXi-CMIPAlign substantially improves CMIP6 projections of climate extremes, providing a robust generative framework for advancing climate risk assessment, mitigation and adaptation.
How to cite: Tie, R., Zhong, X., Shi, Z., Li, H., Chen, B., Liu, J., and Wu, L.: Generative spatiotemporal downscaling model improves projections of climate extremes, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-52, https://doi.org/10.5194/egusphere-egu26-52, 2026.
Global Climate Models (GCMs) are essential tools for climate projections, but their coarse spatial resolution (~100–200 km) and systematic biases limit their direct use for regional impact studies. This limitation is particularly critical for wind-related applications, such as wind energy assessments, which require spatially coherent, multivariate, and physically plausible near-surface wind fields. Classical statistical downscaling and bias correction methods partly address this issue, yet they often fail to preserve spatial structure, inter-variable consistency, and robustness under climate change, especially when applied to high-dimensional climate fields.
Recent advances in generative machine learning offer new opportunities for downscaling and bias correction without relying on explicitly paired low- and high-resolution datasets. Such methods can generate fine-scale, physically consistent fields conditioned on large-scale climate patterns. However, many existing approaches remain difficult to interpret and challenging to deploy in operational climate impact studies.
In this work, we apply SerpentFlow, an interpretable, generative, domain alignment framework, to the multivariate downscaling and bias correction of wind variables from the ACCESS Earth System Model over the French territory at a resolution of approximately 25 km, under the SSP2-4.5 scenario. The framework constructs pseudo low-/high-resolution pairs by explicitly separating large-scale spatial patterns from small-scale variability, aligning large-scale components between model outputs and observations, and learning conditional fine-scale variability via a flow-matching generative model. This approach enables the generation of realistic fine-scale wind fields while preserving physical plausibility and inter-variable correlations.
We evaluate the method on multiple near-surface wind variables, including wind speed, zonal and meridional components, and maximum wind speed, and compare its performance to widely used statistical downscaling and multivariate bias correction methods, such as CDF-t and R2D2. Evaluation metrics include the preservation of spatial structure, inter-variable correlation, extremes, and robustness under future climate conditions. We find that SerpentFlow significantly improves spatial coherence and consistency among wind components compared to baseline methods, while maintaining realistic distributions and extreme events. Ensemble simulations further illustrate the method’s ability to capture stochastic fine-scale variability, an important aspect for climate risk assessment and energy resource studies.
Our results demonstrate that interpretable generative domain adaptation methods can address critical limitations of classical downscaling techniques, providing high-resolution, physically consistent, and multivariate-consistent wind fields suitable for climate impact and energy applications. This work highlights the potential of SerpentFlow as a flexible tool for operational downscaling tasks, capable of adapting to different GCMs, resolutions, and scenarios without requiring paired training data. The framework thus represents a promising avenue for generating reliable, high-resolution climate information to support regional adaptation and wind energy planning.
How to cite: Keisler, J., Oueslati, B., Charantonis, A., Goude, Y., and Monteleoni, C.: Generative Unsupervised Downscaling of Climate Models via Domain Alignment: Application to Wind Fields, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19787, https://doi.org/10.5194/egusphere-egu26-19787, 2026.
Bridging the scale gap between coarse General Circulation Models (GCMs) and high-resolution data, e.g. the type required for hydrological assessment, remains a significant challenge. While dynamic downscaling via Regional Climate Models (RCMs) offers guarantees of physical consistency, its computational cost prohibits creating the large-volume ensembles required for catastrophe risk assessment. This work presents a matured Generative Diffusion Model (DM) framework that achieves high-resolution (10 km) downscaling across Europe with significantly lower computational overhead than similar methods. Crucially, we demonstrate zero-shot transferability by downscaling the 100-member CESM2 Large Ensemble (CESM2-LENS), despite the model being trained exclusively on reanalysis data.
To move beyond traditional pixel-wise metrics, we employ a multi-scale validation strategy: (1) Distributional integrity, recovering extreme precipitation tails; (2) Spatial consistency, using Radially Averaged Log Spectral Density to confirm correct energy distribution from convective scales to synoptic systems; and (3) Temporal coherence, ensuring the chronological sequences required for realistic soil moisture evolution. Finally, we provide an "end-to-end" validation by forcing the Wflow distributed hydrological model. The resulting discharge simulations capture historical extremes across diverse European catchments, proving that the generative output is not merely visually plausible but physically functional. This framework offers a scalable, computationally efficient pathway for generating the massive synthetic event sets required for risk assessment in a non-stationary climate.
How to cite: Lucas, C., Lord, N., Addor, N., Moraga, S., Hoch, J., Marshall, A., and Wing, O.: Generalizable Generative Downscaling: Maintaining Physical Consistency from Reanalysis to GCMs and Hydrological Applications, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19822, https://doi.org/10.5194/egusphere-egu26-19822, 2026.
Regional Climate Models (RCMs) provide high-resolution, physics-based fields, but they face three main limitations. First, they are computationally expensive and hence difficult to scale across scenarios or ensembles. Second, they lack uncertainty quantification. Third, they usually take only coarse data from Earth System Models (ESMs) or reanalysis to predict fields, without assimilating real observations. In response to these problems, neural emulators of RCMs have been developed over different regions.
In this work, we present MAR.ia, a diffusion-based emulator of MAR, an RCM developed at ULiège tailored to Belgium (Doutreloup et al., 2019). Our approach maps coarse atmospheric and surface reanalysis variables (ERA5 at 0.25° and 1° resolution) to key surface variables (temperature, precipitation and wind speed) at the resolution of MAR (5 km). The emulator is conditioned on ERA5 reanalysis every six hours (as the forcing of MAR) in order to give hourly MAR-like fields. We assess the sensitivity of the emulator to the choice of ERA5 fields, identifying the key drivers to reproduce MAR dynamics.
We solve the three main limitations initially stated: we reduce computational costs by several orders of magnitude, we estimate uncertainty by sampling several times for the same coarse inputs (generation of ensembles), and we incorporate observational constraints from ground stations and satellites directly during sampling, while showing competitive metrics, i.e. correlation of ~0.99 for the temperature at 2m.
Future work will attempt to use ESM outputs (weather forecast or CMIP future projections) as context variables instead of reanalysis, enabling both short-term meteorological predictions and long-term climate projections up to 2100, over Belgium.
Doutreloup, S., Wyard, C., Amory, C., Kittel, C., Erpicum, M., and Fettweis, X. (2019). Sensitivity to Convective Schemes on Precipitation Simulated by the Regional Climate Model MAR over Belgium (1987–2017), Atmosphere, 10( 1), 34. https://doi.org/10.3390/atmos10010034.
How to cite: Faulx, E., Peters, S., Fettweis, X., and Louppe, G.: MAR.ia: a diffusion-based emulator for high-resolution climate downscaling over Belgium, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18397, https://doi.org/10.5194/egusphere-egu26-18397, 2026.
Climate impact assessment requires more detailed, sector-specific climate information, especially when impacts depend on crossing specific thresholds, such as heat-stress conditions. Regional climate models (RCMs) can provide such high-resolution climate projections, but systematic biases hinder their direct use. Therefore, bias adjustment (BA) methods are commonly applied in impact studies devoted to heat-stress, which, besides, is a multivariate hazard. Selecting an appropriate BA method for multivariable indices remains challenging due to the need to preserve inter-variable dependence structures and the climate change signal.
This study examines multiple BA methods to generate regional climate projections of two multivariable heat-stress indices—wet-bulb temperature (wbt) and a simplified version of the wet-bulb globe temperature (swbgt)—over southeastern South America (SESA). Both indices rely on temperature and humidity but differ in their sensitivity to these input climate variables. For this assessment, five BA methods were analysed, including trend-preserving and non-trend-preserving techniques as well as univariate and multivariate approaches.
CORDEX and CORDEX-CORE RCM simulations available for SESA driven by three different global climate models were considered, and the MSWX dataset was used as reference. To adjust the indices, an indirect approach was adopted, with the individual input climate variables adjusted prior to index calculation. All methods were trained on austral summer days from the historical period and then applied to RCP8.5 future simulations. Future changes were assessed for the mean and maximum summer values, as well as for two frequency-based metrics using heat-stress thresholds in order to examine the contribution of the RCM and BA method to the overall uncertainty.
Climate change projections obtained from trend-preserving and non-trend-preserving methods considerably differed in magnitude and spatial distributions, with non–trend-preserving approaches typically underestimating the RCMs raw signal, clearly for the mean values. Multivariate methods enhanced the representation of heat-stress indices during training, better capturing the correlation between temperature and humidity, although no added value was identified in the projected delta changes.
Large uncertainties within RCMs raw outputs and BA methods were found in the magnitude of the change signal for the climate input variables, especially for humidity, which were considerably reduced after computing the indices. In particular, the differing sensitivities of the indices to temperature and humidity were highlighted: wbt closely reflected regions with large humidity-related uncertainties, whereas swbgt aligned more closely with the spatial patterns of temperature uncertainties.
This study provides valuable information on the use of BA methods in multivariable impact studies in SESA—a region where fine-scale climate projections remain limited—and underscores the importance of carefully evaluating BA methods prior to climate-impact applications, particularly in a multivariable, climate-change context.
How to cite: Balmaceda-Huarte, R., Casanueva, A., and Bettolli, M. L.: Disentangling the effect of bias adjustment on climate change projections of heat stress in Southeastern South America, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1233, https://doi.org/10.5194/egusphere-egu26-1233, 2026.
With the growing number of statistically downscaled datasets available, it can become difficult for users to choose what to focus on when selecting an ensemble and to understand the impact of this choice. To assist in this task, the authors use a systematic approach to quantify the uncertainty sources of statistically downscaled and bias-adjusted climate simulations. Classical uncertainty partitioning of climate simulations includes internal variability, greenhouse gases scenario and global climate model. Bias adjusted and statistically downscaled datasets descend a level deeper in the cascade of uncertainty. To study this, the authors include two new dimensions: observational reference used in bias-adjustment and bias-adjustment method itself. The fraction of uncertainty associated with each of these five dimensions is calculated for precipitation-based, temperature-based and multivariate indicators. Eastern Canada is used as a case study, focusing on three locations with contrasting climates and observational network densities. This analysis reveals that, while the method is only responsible for a small portion of the variance, the uncertainty associated with the observational reference dataset can play a major role, even becoming the leading source of uncertainty in many cases. This finding underscores the importance of this, often overlooked, dimension in the evaluation of datasets by users and impact modelers. Further, it highlights the ethical responsibility for data providers to clearly communicate the full uncertainty structure of their products.
How to cite: Lavoie, J., Caron, L.-P., Logan, T., Sobie, S., Turcotte, R., Mailhot, E., and Pelletier-Dumont, J.: Partitioning the sources of uncertainty in statistically downscaled and bias-adjusted climate simulations, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5950, https://doi.org/10.5194/egusphere-egu26-5950, 2026.
Reliable projections of Earth’s future climate are an essential source of information to better adapt to the impacts of climate change on societies and natural systems. Climate models provide information on the possible evolution of climate in the coming decades to centuries, however, this information has several limitations such as inadequate resolution to capture the fine-scale features that characterize hydroclimatic conditions at the local scale. Climate model output downscaling aims at partly addressing these limitations.
Here, we present a novel methodology to generate 5 km × 5 km climate information at the European scale based on CMIP6 model output, which not only corrects model biases locally, but also preserves large-scale climate features (spatial correlation) from the original climate model data.
Our approach builds from an existing downscaling technique: Bias-Corrected Constructed Analogues with Quantile Mapping Reordering. Compared to the BCCAQ implementation available in the well-known R package ClimDown, our methodology introduces two major differences:
Identification of Dynamically Coherent and Persistent Weather Regimes: We perform the daily analogue selection only for dynamically coherent and persistent days. This process begins by identifying large-scale circulation patterns. The first 10 principal components (PCs) of daily mean sea level pressure (MSLP) from both the CERRA reanalysis and the GCM are calculated. Then, a multivariate Hidden semi-Markov model (HSMM) is used to detect hidden states (representing meteorological regimes) in the GCM's data over the period 1950–2100. This allows for the identification of persistent blocks of at least five consecutive days characterized by a single dominant weather regime. Blocks shorter than five days, or those without a dominant regime, are excluded from the reordering step.
Targeted Analogue Search and Reordering: For each day within an identified block, the search for historical analogues in the CERRA data is conducted within a window of ±15 days from that calendar day, using a mean squared difference metric on the relevant variable. Finally, a "Schaake Shuffle" reranking of the corresponding Quantile Delta Mapping (QDM) daily outputs is performed within each identified block of continuous days using the identified climate analogues. This ensures the preservation of realistic temporal structure of the weather sequences across the coherent meteorological regimes.
Our downscaling method is calibrated with historical data (1985–2014) from the Copernicus European Regional Reanalysis (CERRA) and this calibration propagates the downscaling into the future for model simulations up to 2099 using the emission scenarios SSP245, SSP370 and SSP585 for the nine climate models and for the variables daily maximum (tasmax), minimum (tasmin), mean (tas) temperature and daily precipitation (pr).
The proposed methodology is portable and potentially applicable to any other region and/or set of input model data as well as an observational reference used to calibrate the model data.
How to cite: Fioravanti, G., Toreti, A., Volpi, D., Hrast-Essenfelder, A., and Acosta-Navarro, J.: A Novel Statistical Downscaling Methodfor Generating High-Resolution ClimateProjections for Europe from CMIP6, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10424, https://doi.org/10.5194/egusphere-egu26-10424, 2026.
Currently, operational weather forecasts rely on physically-based modeling approaches, with Numerical Weather Prediction (NWP) models used to determine atmospheric conditions over the coming hours to days. However, the configuration of NWP models is strongly constrained by computational resources, which notably limits, for instance, their horizontal resolution. Current operational systems typically run at resolutions of around 10 km at the global scale and, at best, around 1 km at the regional scale. A promising alternative to explicitly increasing resolution is statistical downscaling, which consists of learning the relationship between large-scale and fine-scale forecasts. This task, similar to super-resolution, can leverage recent advances in AI for computer vision.
The literature on downscaling approaches for weather and climate prediction is already extensive, with a wide range of AI methods proposed, from standard convolutional neural networks to more advanced generative approaches, including GANs and diffusion models. Generative methods learn a probabilistic representation of the data, which helps avoid the fine-scale blurring commonly encountered in standard AI approaches and naturally enables the generation of ensemble forecasts. However, most existing applications for weather or climate downscaling focus on a limited set of variables or treat each variable independently.
In this work, we develop a diffusion-based downscaling model, termed AROME-DS, to emulate high-resolution forecasts from the French regional model AROME (0.025°) from those of the French global model ARPEGE (0.1°). The model is based on a graph transformer encoder–processor–decoder architecture implemented within the Anemoi framework. It is trained on five years of hourly analyses produced by the French operational services at Météo-France. AROME-DS jointly predicts 70 atmospheric variables, including 11 vertical levels and multiple surface fields such as near-surface temperature, precipitation, and wind gusts, representing a significant increase in variable dimensionality compared to existing AI-based downscaling approaches.
We show that AROME-DS produces realistic high-resolution forecasts and successfully retrieves fine-scale features related to orography. We further investigate how ensemble forecasts obtained by sampling the distribution learned by the diffusion model can be used to represent uncertainty in specific weather situations. Finally, we compare this downscaling approach with an AI-based autoregressive regional NWP model, providing insights onto the best way to leverage AI in operational weather prediction.
How to cite: Lumet, E., Dumont-le-Brazidec, J., Lang, S., Devillers, B., Salas-y-Melia, D., and Raynaud, L.: Diffusion downscaling for regional convective-scale weather prediction, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2566, https://doi.org/10.5194/egusphere-egu26-2566, 2026.
Chemical transport models (CTMs) are essential for assessing air quality and designing mitigation strategies, yet computational constraints typically limit their operational output to coarse resolutions (e.g., 10-15 km over regional domains). These resolutions are often insufficient to capture local pollution hotspots or neighborhood-scale variations required for accurate exposure assessment. In the frame of the Copernicus Atmospheric Monitoring Service (CAMS) National Collaboration Programme (NCP) contract and AIRE SPanish national project, we are investigating the application of deep learning-based super-resolution techniques to downscale atmospheric composition fields while enforcing physical constraints such as mass conservation.
Our research utilizes a large-scale dataset spanning three years (2021-2023) with hourly outputs covering the Iberian Peninsula. We employ the MONARCH chemical transport model to generate 72,000 paired samples, consisting of high-resolution (5 km) ground truth and synthetically coarsened (10 km) inputs for pollutants including NO2, O3, PM10, and PM2.5, alongside high-resolution meteorological fields and anthropogenic emissions (obtained with the HERMES emission module) as auxiliary inputs. We compare the performance of several architectures adapted from computer vision, specifically Convolutional Neural Networks (CNN), Residual Channel Attention Networks (RCAN), and Enhanced Deep Residual Networks (EDSR). A key methodological innovation in our approach is the integration of high-resolution auxiliary data directly into the learning process to guide the reconstruction of pollutant fields. Additionally, we explore architectural modifications such as renormalization layers to enforce hard physical constraints, including mass conservation and non-negativity.
Our results demonstrate that deep learning models significantly outperform traditional deterministic baselines. A primary finding is that the inclusion of high-resolution ancillary data is critical for performance, providing the necessary physical context to recover sharp spatial gradients. We observe that relatively compact models are capable of achieving impressive fidelity; we report Pearson correlation coefficients exceeding 0.988 and normalized Root Mean Square Error (nRMSE) below 20% across all target pollutants. Qualitative inspection confirms these quantitative gains, as the generated high-resolution maps are nearly indistinguishable from the ground-truth simulation fields. However, we also find that increasing model depth introduces training stability challenges, such as gradient explosions, which require careful optimization strategies.
Current efforts are now focused on reducing temporal biases and improving the robustness of the models across different atmospheric perturbations. Future work will extend this framework to higher scaling factors (i.e., downscaling to 2.5 km resolution) and transition from learning on synthetically degraded data to mapping native low-resolution simulation outputs directly to high-resolution targets. The latter is not trivial, as CTMs are not spatially consistent across resolutions due to information loss during the coarsening process. Finally, we aim to explore spatiotemporal architectures to leverage the temporal coherence inherent in atmospheric transport processes.
How to cite: d'Hondt, J. E. and Petetin, H.: Physics-Constrained Deep Learning for Downscaling Atmospheric Chemistry Simulations: The Role of Auxiliary Forcings and Model Architecture, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9829, https://doi.org/10.5194/egusphere-egu26-9829, 2026.
High-resolution emission data are essential for strategic environmental governance and accurate air quality modeling. However, fine-scale (i.e. 1-km) emission assessments remain challenging for traditional bottom-up inventories in Global South countries, including China, due to the lack of unit-level source information. Meanwhile, observation-based emission inversions are often limited in timeliness, spatial resolution, and/or sectoral discrimination. Here, we integrate a fast physics-based inversion framework, PHLET, with big Earth data to derive 1-km-resolution, sector-specific emissions from satellite observations. The resulting new framework, PHLET-BIG, achieves accurate emission positioning and sectoral attribution by incorporating spatial features linked to emission sources extracted from high-resolution Earth data.
Applying PHLET-BIG to China reveals unprecedented fine-scale distributions of NOX emissions and their recent sectoral spatiotemporal evolution during the summers of 2018–2024. Emissions span several orders of magnitude and show a clear decoupling from population density and nighttime light at the 1-km grid scale. While national total NOX emissions declined by 24.6% over this period, pronounced sectoral contrasts persist at individual locations, townships, and counties. PHLET-BIG enables unit-level emission tracking from space, demonstrates consistency with in situ flux observations, and reduces NO2 modeling errors by 20–60%. This framework provides a cost-effective foundation for refined emission control strategies and fine-scale air pollution analyses.
How to cite: Kong, H., Lin, J., and Hu, Y.: PHLET-BIG: 1-km resolution inversion of sectoral emissions based on satellite constrained by big Earth data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1918, https://doi.org/10.5194/egusphere-egu26-1918, 2026.
As climate variability intensifies, extreme weather events are expected to change its frequency and severity, increasing the need for high-resolution meteorological data capable of resolving small-scale processes such as convective storms, urban heat islands, and extreme wind events. The ERA5 reanalysis produced by the European Centre for Medium-Range Weather Forecasts (ECMWF) is widely used for global and regional analyses, but its coarse spatial resolution limits its applicability for fine-scale impact studies. Dynamical downscaling using physical models can bridge this gap but this approach remains computationally expensive. As an alternative, machine learning based models that learn to map coarse data into data produced by physical models offer a computationally inexpensive solution.
Here, we present a multivariate deep learning framework based on a UNet architecture to emulate and downscale key near-surface and convective variables from ERA5 to convection-permitting resolution using limited data. Five low-resolution atmospheric predictors at three pressure levels (850, 700 and 500 hPa), together with five single level variables and a high-resolution elevation map is used as input for the model, which aims to emulate Most Unstable Convective Available Potential Energy (MUCAPE) and downscale 2m temperature and 10m wind components. The model is trained using ERA5 data at 25 km resolution as input and CONUS404, a WRF-based regional hydroclimate reanalysis at 4 km resolution over the contiguous United States, as the target.
Relative to ERA5, the downscaled fields exhibit substantial error reductions, with root-mean-square error improvements of 35.7% for MUCAPE, 20.0% for 2 m temperature, 23.0% for zonal wind, and 20.8% for meridional wind. The model reproduces fine-scale spatial structure, realistic value distributions, and seasonal and temporal variability, and demonstrates skill in representing extreme convective environments, including those associated with hurricanes.
These results highlight the ability of multivariate deep learning to capture complex inter-variable relationships in the atmosphere. In particular, deep learning–based MUCAPE emulation provides a computationally efficient alternative to traditional diagnostic calculations, enabling spatially detailed and readily accessible datasets for severe weather analysis and climate impact studies using a limited set of input variables.
How to cite: Benitez, M., Rodriguez, M., Margalef, T., Panadero, J., and Dutta, O.: Deep Learning Emulation of Convective Instability and Near-Surface Fields from ERA5, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9122, https://doi.org/10.5194/egusphere-egu26-9122, 2026.
Accurate characterization of surface air temperature at the urban scale is relevant for developing effective climate change adaptation and mitigation strategies in the context of global warming. However, reanalysis products such as ERA5-Land provide 2 m air temperature (T2m) at relatively coarse spatial resolutions, limiting their applicability for detailed urban-scale analyses. To address this limitation, this study focuses on the spatial downscaling of ERA5-Land T2m from 0.1° to 0.05° resolution using a deep learning–based approach. A specific type of Convolutional Neural Network (CNN), known as Super Resolution Deep Residual Network (SRDRN), is implemented to enhance the spatial detail of surface air temperature fields. The proposed framework integrates auxiliary variables derived from satellite observations and meteorological reanalysis data to better capture surface–atmosphere interactions and improve model performance. These auxiliary features include the Normalized Difference Vegetation Index (NDVI), Land Surface Temperature (LST), albedo, Normalized Difference Built-up Index (NDBI), as well as meteorological variables such as precipitation, solar radiation, and wind components. Model training and evaluation are performed following a supervised learning approach, with the fine-resolution MERIDA-HRES dataset used as reference data and split into training, validation, and testing subsets. The SRDRN configuration incorporating these multisource auxiliary features outperforms both a previous downscaling experiment based on T2m and baseline methods, including the classical statistical downscaling approach LOcalized Constructed Analog (LOCA) and bilinear interpolation (previous SRDRN: RMSE = 1.4 °C, R² = 0.74). In addition, an evaluation employing the SPHERA dataset at 0.02° spatial resolution further confirms the robustness and spatial consistency of the proposed approach. These results demonstrate that the inclusion of satellite-derived surface data and specific meteorological variables substantially improves the accuracy of downscaled T2m at spatial resolutions closer to the urban scale. By enhancing the spatial resolution of surface air temperature data, this work confirms the potential of deep learning approaches for temperature downscaling and subsequent urban climate analysis. Future work will focus on increasing the spatial resolution to 0.01° and validating the enhanced products against in-situ weather observations to further assess accuracy, robustness, and applicability for urban climate services.
How to cite: Parmeggiani, D., Costanzini, S., Despini, F., Ghermandi, G., and Teggi, S.: Deep learning–based downscaling of ERA5-Land surface air temperature using multisource auxiliary data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18822, https://doi.org/10.5194/egusphere-egu26-18822, 2026.
Deep learning (DL) models have become popular methods to downscale low resolution climate data into high resolution climate projections, with the goal of avoiding the high computational cost associated with dynamical models like Regional Climate Models (RCMs). These DL-based downscaling models when applied in the context of RCMs and their Global Climate Model (GCM) counterparts, are referred to as RCM emulators.Currently, most DL based RCM emulators are single variate, which presents several drawbacks. For example, actual RCM's are multivariate and thus an RCM emulator should be as well. Additionally, a goal of these models is capturing extreme weather events, which are often multivariate as well. As such, this work explores the added value of multivariate emulators by testing four different DL-based RCM emulators (plus a single-variate emulator as baseline) at recreating a daily time series of 2D maps representing the average, maximum and minimum temperature on a given day at surface. All of these models rely on a U-Net based architecture. Notably, two of these DL models are considered to be ''temporal" (one of which implements a ConvLSTM architecture) as they both use multiple days worth of input data to make their predictions. These models are evaluated against a true RCM via several evaluation metrics, including general numerical metrics (RMSE, Correlation, etc.) as well as through real world applications, like the emulators ability to accurately represent future climate or reproduce heatwave events. We also implement a scheme of statistical significance testing via the Kruskal-Wallis method (with Dunn’s as post-hoc). Our results show that the temporal emulators, especially the LSTM model, consistently outperform the other models on a variety of the metrics. The results here support the theory that there is added value in not only making RCM emulators multivariate, but also that temporality improves the emulator's ability to make its predictions.
How to cite: Carty, C.: Multivariate deep-learning based regional climate model emulators and the impact of temporal awareness, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11428, https://doi.org/10.5194/egusphere-egu26-11428, 2026.
Poor air quality poses a major threat to public health globally. Fine particulate matter (PM2.5) is of particular concern due to its ability to penetrate deep into the lungs and enter the cardiovascular system, contributing to respiratory disease, cancer and early mortality. These health impacts underpin the critical need for accurate, high-resolution estimates of population exposure to support effective intervention strategies and safeguard public health.
There are many sources of information detailing air quality, including ground observations, remote sensing and atmospheric models (AM). Ground networks can provide accurate local measurements but are often spatially sparse, while satellite products and AMs often provide good spatial coverage but may lack local detail and may be affected by indirect measurement errors or model misspecification. Data integration modelling techniques can be employed to bring these complimentary data sources together and enable accurate, spatially continuous, high-resolution maps of air quality estimates.
Statistical downscaling approaches are commonly employed for this purpose, but often their high computational cost and limited scalability have motivated the adoption of downscaling through machine learning (ML) methods. However, ML models are traditionally deterministic, not providing explicit quantification of prediction uncertainty which is vital for risk-based decision making. We can address this gap by developing a probabilistic ML downscaling framework based on a Bayesian convolutional neural network (BCNN) where predictive uncertainty deriving from both model structure and random error is quantified using Monte Carlo dropout.
In this study, a BCNN is designed to enhance Copernicus Atmosphere Monitoring Service (CAMS) PM2.5 forecasts from their native 10 x 10 km resolution to 1 km in Western Europe. CAMS spatial data is spatially located with PM2.5 ground observations such that each extracted image corresponds to an observed concentration at a given time and location. The BCNN is trained to learn the relationships between largescale atmospheric patterns and local PM2.5 concentrations, enabling the creation of high-resolution prediction maps even in regions where ground monitoring in limited.
The resulting framework produces spatially detailed, probabilistic PM2.5 estimates at relatively low computational cost compared to traditional statistical downscaling methods. The downscaled pollution data enables improved assessments of population exposure to poor air quality and the identification of pollution hotspots. This approach demonstrates strong potential for broader applications in data-sparse regions and for supporting urban-scale air quality planning.
How to cite: Stella, L., Thomas, M., and Topping, D.: Uncertainty aware Deep Learning for Downscaling Air Quality Concentrations, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7377, https://doi.org/10.5194/egusphere-egu26-7377, 2026.
Abstract
Traditional data assimilation based on numerical models has been utilized for risk assessment and served as a basis for policy decision-making and regulatory establishment. However, data assimilation is constrained by the resolution of the underlying numerical models, presenting limitations in producing high resolution. In this study, we propose a statistical downscaling method to generate 1 km concentration fields for East Asia using a Graph Convolutional Network (GCN) model. The study was conducted in two phases. In Phase 1, the initial concentration fields were derived using the Community Multiscale Air Quality (CMAQ) model, driven by WRF-simulated meteorology and SMOKE-based emission inventories, with further refinement via surface observation data assimilation. In Phase 2, the GCN model was developed to downscale from 27 km to 1 km resolution, using the reanalysis fields from Phase 1, land-use data from WPS, and emission data from EDGAR as input features. The GCN model used semi-supervised learning by masking 70% of surface monitoring stations to separate training and validation data. The model evaluation indicated that the RMSE was 1.28 μg/m³ for PM2.5, 1.5 ppb for O3, and 0.8 ppb for NO2 in China. In the Korean Peninsula, the RMSE was 1.83 μg/m³ for PM2.5, 2.0 ppb for O3, and 1.3 ppb for NO2. The proposed GCN-based statistical downscaling methodology is expected to produce high-quality, high-resolution data that can contribute to risk assessment and policy development.
Acknowledgment
"This research was supported by Particulate Matter Management Speciallized Graduate Program throu the Korea Environmental Industry & Technology Institute(KEITI) funded by the Ministry of Environment(MOE)"
How to cite: Lee, J., Choi, D., Kang, J., and Han, S.: Statistical Downscaling of PM2.5 and Gaseous Pollutants in East Asia Based on Graph Neural Network, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3645, https://doi.org/10.5194/egusphere-egu26-3645, 2026.
Microclimate models are increasingly used to assess the effectiveness of climate change adaptation strategies against future heat stress. These models require high-resolution climate inputs for multiple variables, including precipitation, air temperature, wind, radiation, and humidity. While the highest spatial and temporal resolution climate information is typically provided by regional climate models, particularly convection-permitting models (CPMs), it remains unclear whether CPM outputs still require bias correction across all relevant variables and whether commonly applied methods such as quantile mapping are suitable in this context.
In this study, we evaluated the performance of the convection permitting model, COSMO-CLM, against observations for three Swiss cities, Zurich, Geneva, and Lugano, across six climate variables: precipitation, air temperature, solar radiation, wind speed, surface pressure, and relative humidity. Delta quantile-mapping was applied to bias-correct these variables for a historical period (1998–2009) and a future period (2078–2089), using COSMO-CLM simulations driven by MPI-ESM-LR under the RCP8.5 scenario. Model performance was evaluated using cross-validation for the historical period and by comparing the climate change signal of selected climate indices (e.g., Maximum Daily Air Temperature and Annual Mean Precipitation) between raw and bias-corrected outputs for the future period. Additional analyses examined whether inter-variable correlation structures were preserved after bias-correction and whether diurnal temperature patterns were respected.
The raw COSMO-CLM output exhibits systematic biases across all variables, with particularly pronounced biases in precipitation, temperature, reltaive humidity, and solar radiation. Delta quantile mapping cannnot substantially reducethese biases but can preserve inter-variable correlations. However, climate change signals that are not explicitly represented in the model were incorporated for wind speed, relative humidity, surface pressure, and solar radiation, while climate change signals for precipitation and temperature are not well preserved. In addition, the method exhibits limitations in representing extreme events especially precipitation events above the 99th percentile and can shift the diurnal air temperature distribution. The latter is of particular concern in this context, as mitigation of heat stress during the hottest hours of the day is the primary focus of climate change adaptation against heat. Variable-specific bias-correction approaches may therefore be required; however, such tailoring can complicate the preservation of physically consistent inter-variable correlation structures. In general, it remainschallenging to identify appropriate evaluation metrics for assessing the usefulness and validity of bias-correction techniques when applied across multiple climate variables. Overall, this study presents a multi-variable assessment of the benefits and limitations of quantile mapping for high-resolution climate data used in urban microclimate modeling and climate change adaptation applications.
How to cite: Liu, F., Yin, Y., and Cook, L.: Challenges in Multivariate Bias Correction of Convection-Permitting Climate Models for Urban Microclimate Applications , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18164, https://doi.org/10.5194/egusphere-egu26-18164, 2026.
Univariate bias-correction methods adjust systematic errors in climate model outputs for individual variables but often fail to preserve inter-variable dependence, resulting in physically inconsistent multivariate projections. Multivariate bias-correction (MBC) methods address this limitation but are commonly applied independently at each location, thereby neglecting spatial dependence. Moreover, temporal dependencies are rarely modeled explicitly. Preserving spatiotemporal consistency is, however, essential for realistic climate dynamics and reliable regional impact assessments.
We propose a novel MBC framework that jointly accounts for inter-variable, spatial, and temporal dependence. The spatiotemporal structure is addressed by decomposing each time series using generalized additive models (GAMs) to remove deterministic components such as seasonality and spatial gradients. The resulting stochastic components are transformed via probability integral transforms into approximately independent and identically distributed variables, suitable for dependence modeling with vine copulas.
To construct a joint distribution across multiple variables and locations, we introduce CUVEE (Copulas Under Vine Extending Environment), a hierarchical vine-based merging strategy. CUVEE combines two dependence levels: (i) spatial dependence across locations modeled separately for each variable, and (ii) inter-variable dependence modeled at a selected reference location, which links the spatial models into a coherent multivariate and spatial structure. This approach enables flexible dependence modeling while remaining computationally tractable for regional applications.
We apply the proposed method to EURO-CORDEX simulations over the Swiss canton of Vaud, using gridded MeteoSwiss observations and ERA5 reanalysis data as reference. Results show substantial improvements in preserving inter-variable, temporal, and spatial dependence compared to standard quantile mapping and conventional MBC approaches, highlighting the potential of the method for physically consistent multivariate bias correction.
How to cite: Meier, T., Chavez-Demoulin, V., Koch, E., and Vatter, T.: Ensuring spatiotemporal consistency in multivariate bias correction for climate projections using hierarchical vine copulas and GAMs, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3121, https://doi.org/10.5194/egusphere-egu26-3121, 2026.
Though the underlying fields associated with vector-valued environmental data are continuous, observations themselves are discrete. For example, climate models typically output grid-based representations of wind fields or ocean currents, and these are often downscaled to a discrete set of points. By treating the area of interest as a two-dimensional manifold that can be represented as a triangular mesh and embedded in Euclidean space, this work shows that discrete intrinsic Gaussian processes for vector-valued data can be developed from discrete differential operators defined with respect to the mesh. These Gaussian processes account for the geometry and curvature of the manifold whilst also providing a flexible and practical formulation that can be readily applied to any two-dimensional mesh. These models can capture harmonic flows, incorporate boundary conditions, and model non-stationary data and can be applied to downscaling stationary and non-stationary gridded wind data on the globe, and to inference of ocean currents from sparse observations in bounded domains.
How to cite: Gillan, M., Siegert, S., and Youngman, B.: Discrete Gaussian Vector Fields on Meshes and their Application to Downscaling, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-549, https://doi.org/10.5194/egusphere-egu26-549, 2026.
Cities contain a significant proportion of the global population. Because of their unique vulnerabilities to climate-related phenomena, such as the Urban Heat Island effect, understanding urban microclimates is essential to the durable safety and well-being of residents. However, global and regional climate models operate at scales too coarse to capture urban-scale processes. Accurately modeling urban microclimates requires resolving fine-scale details, such as the geometry and arrangement of buildings. Such high-resolution simulations entail substantial computational costs, which severely limit their applicability. Because of this, at this time, real-time prediction and design optimization problems remain mostly inaccessible. Therefore, there is a need for computationally efficient urban microclimate models.
The DANTE project aims to address this need by applying model order reduction techniques to high-resolution urban-scale simulations. Resulting models must undergo a rigorous validation process before any application is possible, to ensure accuracy and reliability for real-world applications. This validation process requires urban-scale ground truth data, which is not directly available. Instead, lower-resolution data must be downscaled to urban scale. As a result, downscaling is a critical part of developing reliable urban microclimate models.
The goal of our work is to construct a downscaling framework adapted to the context of weather data, leveraging regional model data, weather station measurements, as well as physical knowledge. In this context, pre-existing high-resolution data is very limited, rendering purely statistical downscaling approaches unsuitable. Since no models - other than those intended for evaluation - are available at the target scale, dynamical downscaling methods are also inadapted. Finally, the inhomogeneity of relevant scales, and the need to integrate data at arbitrary locations requires the use of irregular, variable grids.
A promising approach is to use Physics-Informed Neural Networks (PINNs). PINNs incorporate physical constraints into the learning process by including partial differential equation residuals into the loss function. By using networks that take coordinates as input and output the local system state, a fitted model can be evaluated at arbitrary locations, providing a way to downscale without need for a structured grid.
A major limitation of PINNs is their lack of robustness during training, as convergence can be difficult to achieve reliably. A contributing factor is that different loss terms can have wildly different scales and convergence rates, which can hinder optimization. Previous studies have explored strategies to make convergence more likely, but such results do not always generalize are are typically task and problem-specific.
In this work, we investigate the applicability of PINNs to the downscaling of weather data, formulated as a fluid dynamics problem on unstructured meshes. We assess the performance levels that can be achieved and examine the methodological choices that influence them, including network architecture, collocation point density, loss-term weighting strategies, data preprocessing, and training protocols. We also analyse the associated difficulties, computational costs, and practical requirements, and quantify the added value of the inclusion of physical knowledge.
How to cite: Malhomme, N. and Stabile, G.: Downscaling using Physics Informed Neural Networks for model evaluation at urban scale, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10207, https://doi.org/10.5194/egusphere-egu26-10207, 2026.
In this study, we developed SOFT CUBE, a scenario-based method to rapidly generate building-resolving three-dimensional wind and air temperature fields by combining a precomputed CFD database with operational mesoscale forecasts. For this, we constructed a CFD scenario library for a 2 km × 2 km urban domain by varying inflow wind speed and direction and surface thermal forcing, and supplemented it with auxiliary cases to represent background vertical wind structure and temperature stratification. Then, for each forecast time, we selected and linearly interpolated scenarios consistent with LDAPS boundary-layer conditions and synthesized the full 3D fields by performing layer-by-layer synthesis across the vertical levels. For validation of the developed method, we used LDAPS forecasts as background forcing and compared SOFT CUBE outputs with LDAPS-driven CFD simulations and observations from four urban stations during July–December 2021. The results showed that SOFT CUBE substantially improved near-surface wind-speed estimates compared with LDAPS, reduced air-temperature errors on average, and reproduced spatial patterns similar to those from the coupled LDAPS–CFD model for most cases. Finally, SOFT CUBE reduced the per-case runtime from 141 min for coupled CFD simulations to 3 min, supporting operational-scale high-resolution urban meteorological field production.
How to cite: Wang, J.-W., Lee, S.-H., and Kim, J.-J.: Development of SOFT CUBE: A synthesis framework for urban 3D flow and air temperature using precomputed CFD scenarios and mesoscale forecasts, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3656, https://doi.org/10.5194/egusphere-egu26-3656, 2026.
Machine learning-based weather forecasting models now surpass state-of-the-art numerical weather prediction systems, but training and operating these models at high spatial resolution remains computationally expensive. We present a modular framework that decouples forecasting from spatial resolution by applying learned generative super-resolution as a post-processing step to coarse-resolution forecast trajectories. We formulate super-resolution as a stochastic inverse problem, using a residual formulation to preserve large-scale structure while reconstructing unresolved variability. The model is trained with flow matching exclusively on reanalysis data and is applied to global medium-range forecasts. We evaluate (i) design consistency by re-coarsening super-resolved forecasts and comparing them to the original coarse trajectories, and (ii) high-resolution forecast quality using standard ensemble verification metrics and spectral diagnostics. Results show that super-resolution preserves large-scale structure and variance after re-coarsening, introduces physically consistent small-scale variability, and achieves competitive probabilistic forecast skill at 0.25° resolution relative to an operational ensemble baseline, while requiring only a modest additional training cost compared with end-to-end high-resolution forecasting.
How to cite: Delefosse, A., Charantonis, A., and Béréziat, D.: Super-Resolving Coarse-Resolution Weather Forecasts with Flow Matching, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-22472, https://doi.org/10.5194/egusphere-egu26-22472, 2026.
We present a novel downscaling methodology that addresses the critical challenge of spatial heterogeneity in coarse-scale Gravity Recovery and Climate Experiment (GRACE) and its Follow-On (GRACE-FO) data. Accurately capturing this heterogeneity is essential for local-scale hydrological applications. While machine learning approaches such as the global random forest (GRF) model have been used, the aspatial nature of the GRF model limits its ability to capture spatial heterogeneity when downscaling GRACE (-FO) data. To overcome this, we propose a Geographically Weighted Random Forest (GWRF) model, which integrates spatial weighting into the GRF algorithm to downscale groundwater storage anomalies (GWSAs) to 0.1° resolution over the North China Plain (2003-2025). The added value of this approach is rigorously quantified through benchmarking. We found that the GWRF model outperforms the GRF model, increasing R2 from 0.957 (GRF: training) and 0.73 (GRF: testing) to 0.999 (GWRF: training) and 0.897 (GWRF: testing). The high-resolution GWSAs output exhibits a strong correlation (r = 80) with independent in-situ groundwater observational measurements, thereby enhancing its credibility. The downscaled GWSAs data provide a tangible application, revealing significant groundwater depletion in the Piedmont Plain (PP: -13.42 mm/yr), Yellow River Plain (YRP: -13.25 mm/yr), Hai River Plain (HRP: -12.68), and a moderate depletion in the Coastal Plain (CP: 5.98 mm/yr) sub-regions of NCP. Using a two-stage Generalized Additive Model (GAM), we quantitatively attribute 69-83% of the GWSAs decline to anthropogenic drivers (primarily cropland expansion, NDVI, and population growth) and 7-12% to climatic factors (downward shortwave radiation, precipitation, and sea surface temperature). This work advances downscaling techniques by demonstrating how geographically-aware machine learning can unlock finer-scale insights from GRACE (-FO) satellite data, providing a valuable tool for climate impact assessments and water resource management.
How to cite: Ali, S., Chen, Q., and Wang, F.: Integrating Spatial Weights into Random Forest to Overcome Aspatial Limitations in GRACE data Downscaling: Tracking Groundwater Depletion in the North China Plain, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8725, https://doi.org/10.5194/egusphere-egu26-8725, 2026.
Kilometre-scale weather and climate datasets are invaluable for quantifying, forecasting and projecting hazards in areas of complex topography, such as the Alps. However, producing such datasets using traditional numerical weather prediction (NWP) models is becoming prohibitively expensive, particularly for climate-timescale simulations and large ensembles. Probabilistic generative downscaling offers a potential alternative, as it learns the conditional mapping from coarse global drivers to kilometre-scale regional fields.
Here, we evaluate a modified conditional generative correction–diffusion model (CorrDiff) for downscaling the ERA5 and IFS-ENS datasets over the Greater Alpine Region. The modified CorrDiff model was trained using a 20-year, 1-km resolution dataset produced with the ICON numerical model, with precipitation constrained to Swiss radar observations using a latent-heat nudging scheme. This setup allows us to make a direct comparison with MeteoSwiss' operational NWP ensemble.
Verification against observations and gridded products reveals that CorrDiff achieves competitive performance following substantial targeted adaptations to the model. Although not explicitly encoded in the loss function, the adapted model reproduces emergent climatological indices, including the diurnal cycle of land precipitation and exceedance probabilities for heavy precipitation. It also captures the spatial patterns of consecutive dry and wet days, as well as prevailing wind direction and directional variability.
How to cite: Leutwyler, D., Stamenkovic, P., Arpagaus, M., McGlohon, M., Mishra, S., Lapillonne, X., Schemm, S., and Fuhrer, O.: Generative Diffusion Downscaling for the Alps: Benchmarking CorrDiff against MeteoSwiss Operational NWP Ensemble, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1470, https://doi.org/10.5194/egusphere-egu26-1470, 2026.
Deterministic deep learning models used for climate downscaling often exhibit spectral collapse, resulting in overly smoothed fields that underestimate extreme events. Although Generative Adversarial Networks (GANs) can preserve high-frequency details, their training instability limits the reliability of ensemble generation. Denoising Diffusion Probabilistic Models (DDPMs) offer a solution to both of these problems. They sample from learned probability distributions through iterative denoising, which introduces inherent randomness. This allows each inference to produce statistically different but physically plausible results, a feature that is essential for quantifying uncertainty in climate projections. This study presents the first systematic analysis of ensemble convergence for DDPM-based climate downscaling at a 10× spatial resolution (1.0° → 0.1°). We evaluated configurations with ensemble sizes ranging from 2 to 50 members, focusing on 30 extreme temperature events. Using the multi-modal sampling capabilities of DDPMs, achieved through different random initializations in the reverse diffusion process, we assessed the trade-offs between accuracy, uncertainty, and computational cost. This was done using a set of metrics: RMSE, MAE, Pearson R, SSIM, and PSNR. The research results demonstrate significant convergence trends: (1) ensemble mean predictions exhibit rapid saturation, with 5-member configurations attaining 96–98% of peak performance (RMSE: 0.459°C compared to 0.453°C for 25 members); (2) spatial uncertainty estimates (0.165–0.170°C) stabilize at 5–10 members, with only minor enhancements of less than 1% beyond this point; (3) computational costs increase substantially, a 50-member ensembles necessitate 35 hours, whereas 5-member ensembles require only 4 hours, indicating an 89% reduction in cost with minimal compromise in accuracy. The optimal range of 5–10 members provides strong uncertainty constraints and enables operational scalability in continental-scale applications. In contrast to deterministic models that provide only point estimates or GANs prone to mode collapse, DDPMs' generative sampling inherently quantifies prediction confidence via ensemble spread, thereby encompassing both epistemic model uncertainty and aleatoric variability. This research provides actionable guidance for uncertainty-aware climate downscaling, demonstrating that small DDPM ensembles effectively produce probabilistic projections, which are crucial for evaluating climate risk.
How to cite: Gupta, V., Jha, S. K., Sharma, P. J., Mishra, A., and Joshi, S.: Uncertainty Quantification in Generative Climate Downscaling: A Multi-Ensemble DDPM Analysis, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13665, https://doi.org/10.5194/egusphere-egu26-13665, 2026.
