Topics include (but are not limited to):
- Subgrid-scale parameterization via machine learning (for example those related to air-sea & land-atmosphere interactions)
- Emulators of physical processes, model components, or whole weather and climate models (including end-to-end learning)
- Hybrid ML-physics modelling frameworks
- Foundation Models
- Reinforcement learning (such as for ensuring physical consistency, stability, optimising model behaviour and improving time-series modelling)
- Physics-informed neural networks, neural operators, and differentiable programming
- Verification of data-driven models (including AI forecasting)
- Physical behaviour, encoding and analysis of AI parametrisations, emulators and whole models (such as through feature-based evaluation/conditional vs unconditional evaluation)
- Calibration and parameter optimization using ML
- Coupling of ML models with physical models
- Cross-domain applications (atmosphere, ocean, cryosphere, land).
This session provides a critical overview of current progress and emerging directions in the application of ML across parametrisations, emulation and hybrid modelling.
Orals: Thu, 7 May, 14:00–08:50 | Room -2.15
The current generation of hybrid machine learning and physics-informed machine learning is often limited by the missing availability of comprehensive differentiable models: either strongly simplified models have to be used or machine learning (ML) can’t be integrated natively into process-based models and must be trained separately. Here, we present the ongoing development of SpeedyWeather.jl: A general circulation model that’s differentiable, GPU-capable and ready for ML simulations. SpeedyWeather.jl is a spectral atmospheric GCM with a primitive equation core on flexible grid implementations from Gaussian to HEALPix. It contains simple yet interactive representations of ocean, land and sea ice for coupled climate simulations. With a user interface made for modularity and interactivity, it’s ideally suited as a framework for hybrid atmospheric models. For example, new parameterizations can be defined without any lines of code for GPU or differentiability specifics, yet integrate seamlessly into those. We document the process to achieve differentiability of our model using the general purpose automatic differentiation library Enzyme, problems we encountered and solutions we found. We demonstrate the differentiability with a sensitivity analysis of our model, initial developments of data-driven parameterizations, and give an outlook on the development of differentiable Earth system models.
How to cite: Gelbrecht, M., Klöwer, M., Groenke, B., and Boers, N.: Differentiable Atmospheric Modelling with SpeedyWeather.jl , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1927, https://doi.org/10.5194/egusphere-egu26-1927, 2026.
Weather forecasting is critical for a range of human activities including transportation, agriculture, industry, as well as the safety of the general public. Over the last two years, machine learning models have shown that they have the potential to transform the complex weather prediction pipeline, but current approaches still rely on numerical weather prediction (NWP) systems, limiting forecast speed and accuracy. In this talk, I will give some of the background on these developments. I will then introduce a machine learning model which can replace the entire operational NWP pipeline. Aardvark Weather, an end-to-end data-driven weather prediction system, ingests raw observations and outputs global gridded forecasts and local station forecasts. Further, it can be optimised end-to-end to maximise performance over quantities of interest. I will show that the system outperforms an operational NWP baseline for multiple variables and lead times for gridded and station forecasts. These forecasts are produced with a remarkably simple neural process model using just 8% of the input data and three orders of magnitude less compute than existing NWP and hybrid AI-NWP methods. We anticipate that Aardvark Weather will be the starting point for a new generation of end-to-end machine learning models for medium-range forecasting.
How to cite: Turner, R.: Aardvark weather: end-to-end data-driven weather forecasting, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4536, https://doi.org/10.5194/egusphere-egu26-4536, 2026.
In light of the urgent need to accelerate measures for the adaptation to and mitigation of climate change, accurate Earth system models are more important than ever, for technology assessment and the identification of the most effective climate protection strategies. Global climate models have successfully projected consequences of different future scenarios, but the spread in projections remains large, with subgrid-scale parametrizations being the main origin of these uncertainties. Recently, machine learning-based hybrid models have successfully enhanced parametrizations - their directly data-driven structure can more effectively capture the empirical aspects of the parametrizations. Especially for the more complex parametrizations, such as microphysics or turbulence, which we study here, quantum computing could bring decisive further improvements as a part of hybrid models. Atmospheric turbulence strongly affects weather and climate because it determines the rates of exchange of heat, moisture, and momentum between the earth surface and the atmosphere. However, due to the chaotic nature of turbulence and the wide range of turbulent regimes in the atmospheric boundary layer from deep convection to nearly laminar stable conditions, it is notoriously hard to predict and model.
Here, we develop a prototype of a quantum machine learning-based subgrid-scale parametrization for the vertical temperature flux caused by atmospheric turbulence based on semi-idealized Large-Eddy-Simulations. We run experiments with dry convective boundary layers with the PALM model system. The setups span an 8x8 km2 domain with a resolution of 10 m and horizontal periodic boundary conditions and an imposed surface heat flux, combining runs with different surface heat fluxes and geostrophic winds in our training data set. We train quantum and classical neural networks with different architectures, and find that quantum models based on parametrized circuits with just 2 or 3 qubits achieve accuracies similar to classical models with the same number of trainable parameters, highlighting the possibility to use quantum computing for parametrizations in the near future. In contrast, the Smagorinsky closure deviates strongly from the true flux in this setup. Our quantum and classical cell-based models both generalize well to data from PALM runs with unseen parameters close to the seen range. We further analyze the feature importance in quantum and classical models and find that most of our quantum models show better stability of the Shapley values with respect to varying the random initial conditions of the training runs. Since the number of required qubits to capture the idealized setting is low, it is promising to extend our model to more complex settings with realistic topography and varied weather conditions in the future, e.g. by using ICON boundary conditions in PALM, opening the possibility to exploit quantum advantages anticipated by the more stable interpretability of our prototype models.
How to cite: Dogra, L., Klamt, J., Eyring, V., and Schwabe, M.: Quantum machine learning-based parametrization for boundary layer turbulence, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5888, https://doi.org/10.5194/egusphere-egu26-5888, 2026.
Over the past two years, ECMWF has rapidly developed and operationalised two machine-learned forecasting systems: AIFS Single, a deterministic model, and AIFS-ENS, a fully probabilistic forecasting system. Both systems are trained on ERA5 reanalysis data and further fine-tuned using operational IFS analyses. In this talk, we briefly introduce the AIFS framework and present ongoing research aimed at further improving its forecast skill.
Current efforts are driven by several research directions, including increasing spatial resolution and incorporating observational data. The current AIFS models operate at the native ERA5 resolution of approximately 30km. While higher resolutions could significantly improve forecast skill in surface variables, available datasets, such as the operational IFS analysis at 9km, are only available for a limited number of years. To address this, we explore a cross-resolution fine-tuning strategy in which AIFS is first pretrained on ERA5 at coarse resolution and subsequently fine-tuned on six years of recent operational IFS analyses at 9 km. We present promising early results showing that this approach enables stable fine-tuning down to 9 km and leads to significant gains in surface forecast skill.
A second research direction investigates the use of alternative datasets to improve total precipitation forecasts. ERA5 is known to exhibit deficiencies in the representation of precipitation, particularly in the tropics. We therefore fine-tune AIFS using the Integrated Multi-satellitE Retrievals for GPM (IMERG) dataset, which has been shown to better capture precipitation characteristics in this region. Early results indicate that incorporating IMERG data can significantly improve total precipitation forecast skill in AIFS, with the largest benefits observed, as expected, in tropical regions.
How to cite: Moldovan, G., Prieto Nemesio, A., Pinnington, E., Lang, S., Polster, J., O'Brien, C., Santa Cruz, M., Alexe, M., Cook, H., Forbes, R., and Chantry, M.: Improving AIFS Forecast Skill through Fine-Tuning across Spatial Resolutions and Datasets, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6950, https://doi.org/10.5194/egusphere-egu26-6950, 2026.
Machine learning is increasingly used to analyze, predict, and interpret Earth-system behavior. Here we synthesize AI4OCEANS research to identify practical, transferable lessons for developing ML methods that remain robust when applied to real Earth-system data and are evaluated across regions, scales, and event types. We present methodological advances and common pitfalls encountered when building ML workflows for prediction and diagnosis across oceanic and atmospheric contexts. Emphasis is placed on (i) constructing physically meaningful predictors and representations that generalize beyond a single region or period, (ii) designing evaluation strategies that reflect scientific and decision-relevant objectives (including event- and regime-aware metrics where appropriate), and (iii) quantifying uncertainty and interpretability in ways that support scientific insight rather than purely empirical skill. We further discuss when hybrid strategies—combining statistical learning with physical constraints or dynamical context—improve robustness in specific applications. By framing diverse studies through shared methodological questions across geophysical systems (from coastal ocean change through high-impact atmospheric events and into bycatch threats to marine wildlife), the produced frameworks provide guidance for ML development that is directly relevant to Earth-system modelling and prediction, particularly for variability, extremes, and environmental risks and impacts under anthropogenic influences.
How to cite: Nieves, V.: Transferable Machine-Learning Practices for Earth-System Prediction and Diagnosis, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7654, https://doi.org/10.5194/egusphere-egu26-7654, 2026.
Recent advances in data-driven weather forecasting have demonstrated skill at medium-range lead times, yet often rely on extremely large models, massive training datasets, and substantial computational resources. In this talk, we present a novel quantum-inspired machine learning (QIML) approach for sub-seasonal weather forecasting that prioritizes computational efficiency and dynamical stability, while retaining competitive predictive skill.
First, by using quantum circuits ansätze and entanglement, we design scalable quantum reservoir computing models. The implemented model is parallelizable across multiple GPUs and runs on classical hardware in a quantum-inspired setting. Second, we train our model on ERA-5 reanalysis data for 2m temperature, multiple pressure levels, and precipitation on a global grid. We show that, using an encoder-decoder architecture in conjunction with the proposed QIML model, we demonstrate forecasts of key atmospheric variables up to 45 days ahead. Third, we benchmark our model against state-of-the-art AI for weather forecasting methods and show that the QIML model can produce reliable forecasts for weather and climate extremes, while requiring 10-50X less compute. Fourth, replacing conventional neural architectures with quantum-inspired circuit dynamics enables enhanced physical interpretability and consistency, as the model state evolves according to Schrödinger-type dynamics. We further analyze the learned latent representations using operator-theoretic and spectral tools, revealing coherent structures associated with dominant atmospheric modes.
This work proposed a novel direction to the growing ecosystem of hybrid ML physics approaches by offering a new class of lightweight, stable, and scalable forecasting models that can be deployed efficiently for localized and resource-constrained settings.
How to cite: Ahmed, O., Qazi, S. A., and Magri, L.: Quantum-inspired machine learning for efficient and reliable weather forecasting , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21434, https://doi.org/10.5194/egusphere-egu26-21434, 2026.
Accurate short-term prediction of clouds and precipitation is critical for severe weather warnings, aviation safety, and renewable energy operations. Traditional mesoscale numerical weather prediction models require significant modeling expertise and computational infrastructure. We introduce Stormscope, a family of transformer-based generative diffusion models trained directly on high-resolution, multi-band geostationary satellite imagery and ground-based radar over the Continental United States. Stormscope produces forecasts at a temporal resolution as high as 10 min and 6-km spatial resolution. Geostationary satellites and ground-based radar provide high-resolution, high-frequency observations essential for characterizing the evolving structure of the mesoscale atmosphere. Evaluated against extrapolation methods and operational mesoscale NWP models such as HRRR, Stormscope achieves leading performance on standard verification metrics including Fractions Skill Score and Continuous Ranked Probability Score across forecast horizons from 1 to 6 hours. By operating in native observation space, Stormscope establishes a new paradigm for AI-driven nowcasting with direct applicability to operational forecasting workflows. The approach is highly extensible, with demonstrated computational scaling to larger domains and higher resolutions. Critically, because Stormscope relies solely on globally ubiquitous satellite observations and radar where available, it offers a pathway to extend skillful mesoscale forecasting to oceanic regions and countries without existing strong operational mesoscale modeling programs.
How to cite: Pathak, J., Abbas, M. S., Harrington, P., Hu, Z., Brenowitz, N., Ravuri, S., Durran, D., Adams, C., Hennigh, O., Geneva, N., Leinonen, J., Carpentieri, A., and Pritchard, M.: Storm-scale forecasting from observations, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15271, https://doi.org/10.5194/egusphere-egu26-15271, 2026.
Machine learning is reshaping the entire climate-modelling chain, from climate model development to the study of climate extreme events and their impacts. One of the key drivers of this revolution is the availability of datasets that are sufficiently large for training and validation. For climate extreme events, however, this requirement poses seemingly insurmountable challenges: we need to assess the impacts of unprecedented events for which historical data are too scarce; we must rely on models, yet simulating extremely rare events with them is prohibitively expensive; and any statistical approach, including machine learning, suffers from a severe lack-of-data problem.
Here, we argue that the only viable path forward is to integrate machine learning directly into the data-generation process, in close interaction with state-of-the-art physics-based climate models and observational datasets.
The first building block of our approach is the development of state-of-the-art climate model emulators. AI models trained on historical reanalyses to emulate the dynamics of the global atmosphere have demonstrated both high forecast skill and drastically reduced computational costs. Some of these AI emulators can generate stable trajectories spanning multiple decades, which, combined with their affordability, has the potential to significantly reduce uncertainties related to extreme weather. However, it remains impossible to directly validate whether AI emulators can reliably estimate the risk of extreme events with return times exceeding the historical record. To address this issue, we develop a methodology based on state-of-the-art architectures, with the explicit requirement that emulators exhibit extremely long-term stability, high fidelity, and a faithful reproduction of the stationary statistics of the climate model.
In a first-of-its-kind experiment, we simulate 100,000 years of a stationary climate using PlaSim, a coarse-resolution general circulation model. We then train a set of stable AI emulators using only 100 years of data, and compare the return times of extreme heat waves over Western Europe and the Pacific Northwest, as well as severe precipitation events over the Tropics.
The second building block of our approach consists of rare-event simulation techniques that reduce by several orders of magnitude the computational cost of sampling extremely rare events with CMIP-class climate models. The third building block is the blending of historical observations with CMIP model output within a Bayesian framework to estimate the
probability of extremely rare events constrained by observations. In this talk, we also briefly discuss the second and third building blocks and their connections to the first within a comprehensive, integrated framework.
How to cite: Lancelin, A., Bouchet, F., Wikner, A., Hassanzadeh, P., Dubus, L., and Werner, P.: Solving the lack of data issue for machine learning for rare climate events, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20131, https://doi.org/10.5194/egusphere-egu26-20131, 2026.
In the context of increasing interest in machine learning-based emulators to overcome the computational cost and limited scenario coverage of Earth System Models (ESMs), some key challenges remain, such as capturing internal variability, handling non-stationarity, and realistically representing compound and extreme events — especially at high spatial and temporal resolution.
We present a probabilistic multivariate emulator of climate variables based on a flow matching model trained on the CESM2 Large Ensemble (Danabasoglu et al., 2020) under the SSP3-7.0 scenario. Our approach leverages a flow matching framework (Lipman et al., 2022) to reproduce the spatiotemporal variability of temperature and precipitations on a monthly timescale. The model is conditioned on greenhouse gas concentrations and we evaluate the capability of the model to generate physically consistent climate fields and to capture the full ensemble spread of the original ESM, including tail behavior and potential extreme events. To ensure a better reproduction of observed climatological variability, the flow matching model is fine-tuned on ERA5 reanalyses (Hersbach et al., 2020). This should enable the emulator to act as a stochastic weather generator of plausible climate states under GHG forcing trajectories, accounting for the non-stationarity introduced by anthropogenic climate change, and allowing for the assessment of rare or compound extreme events within the generated ensemble. Our results assess the model’s ability to reproduce ensemble-scale statistics, cross-variable dependencies, and evolving climate distributions across time.
How to cite: Launeau, V., Vrac, M., Lemordant, L., and Gentine, P.: Emulating climate variability and extremes with a multivariate flow matching model trained on CESM2 and finetuned on ERA5, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16454, https://doi.org/10.5194/egusphere-egu26-16454, 2026.
High-resolution imagery is limited by sensor technology, atmospheric effects, and acquisition costs. This is a well-known challenge in satellite remote sensing, but it also applies to ground-level imaging with handheld devices such as smartphones. Super-resolution seeks to overcome these limitations by enhancing image resolution algorithmically. Single-image super-resolution, however, is an ill-posed inverse problem and therefore depends on strong priors, typically learned from high-resolution training data or imposed through auxiliary information such as high-resolution guidance from another modality. While these methods often produce visually appealing results, they are prone to hallucinating structures that do not reflect the true scene content.
Multi-image super-resolution (MISR) addresses this issue by exploiting multiple low-resolution views of the same scene that are captured with sub-pixel shifts. In this work, we introduce SuperF, a test-time optimization approach for MISR based on coordinate-based neural networks, also known as neural fields. By representing images as continuous signals using implicit neural representations (INRs), neural fields are well suited for reconstructing high-resolution images from multiple aligned observations. The central idea of SuperF is to share a single INR across all low-resolution frames while jointly optimizing the image representation and the sub-pixel alignment between frames.
Compared to prior INR-based approaches adapted from burst fusion and layer separation, SuperF directly parameterizes the sub-pixel alignment using optimizable affine transformation parameters and performs the optimization on a super-sampled coordinate grid corresponding to the target output resolution. We evaluate the proposed method on simulated bursts of satellite imagery as well as on ground-level images captured with handheld cameras, and observe consistent improvements for upsampling factors of up to 8. A key advantage of SuperF is that it operates entirely at test time and does not rely on any high-resolution training data.
How to cite: Jyhne, S., Igel, C., Goodwin, M., Andersen, P.-A., Belongie, S., and Lang, N.: Super-Resolving Any Place on Earth - Implicit Neural Representations for Sentinel-2 Time Series, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-14956, https://doi.org/10.5194/egusphere-egu26-14956, 2026.
We introduce FloeNet, a data-driven emulator architecture trained on the Geophysical Fluid Dynamics Laboratory (GFDL) global sea ice model, SIS2. FloeNet is an auto-regressive graph neural network (GNN) which marks a step forward in sea ice emulation as the first model to dynamically evolve the state of sea ice and snow-on-sea-ice by mass and area budget decompositions. Specifically, FloeNet receives mechanical and thermodynamic forcing inputs from the atmosphere and ocean, and predicts ice and snow mass tendencies due to growth, melt, and advection. This yields a mass-conservative and interpretable model, as timestep-to-timestep changes in sea ice area and mass can now be attributed to each term in their respective budget.
Sea ice is often seen as a barometer for climate change. It is therefore crucial that data-driven sea ice models show an accurate response to different climate forcings. To this end, we show how FloeNet successfully reproduces sea ice trends and variability of pre-industrial and 1% CO2 climates, despite being trained only on a present-day climate; FloeNet also reaches globally ice-free conditions under 1% CO2 forcing, with consistent timing to that of the original numerical model. In summary, FloeNet is a fast global sea ice emulator, taking 4.75 hours to generate a 140-year simulation on 1 GPU. It is also stable and accurate, reproducing critical features of long-term sea ice evolution under different forcings. We expect that FloeNet will substantially improve the representation of atmosphere-ice-ocean interactions in existing climate emulators.
How to cite: Gregory, W., Bushuk, M., Duncan, J., Wu, E., Subel, A., Clark, S., McGibbon, J., Henn, B., Arcomano, T., Perkins, W. A., Kwa, A., Watt-Meyer, O., Adcroft, A., Bretheron, C., and Zanna, L.: Towards a mass-conservative global sea ice emulator that generalizes across climates, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11736, https://doi.org/10.5194/egusphere-egu26-11736, 2026.
The Peruvian Andes contain over 70% of the world's tropical glaciers, which are vital for regional water security and are rapidly destabilising due to climate change. Current large-scale projections often lack the spatial resolution required for localised glacial melt modelling or rely on climate reanalysis products that are too coarse in rugged terrain. This study introduces a unified framework that combines high-resolution remote sensing (Sentinel-2, Landsat-8) with machine learning to characterise, monitor, and forecast glacial evolution across Peru from 2016 to 2100.
We propose a machine-learning modelling approach that addresses both the where and when of glacial retreat. We developed a spatial Random Forest classifier to generate country-wide melt vulnerability maps. Ensemble analysis of driving parameters reveals that "distance-to-edge" and topographic factors (elevation, slope) are significantly stronger predictors of melt spatiality than available coarse-resolution temperature and precipitation datasets. Our spatial model achieves a 74.9% overlap accuracy between simulated and observed melt (2016–2023), nearly doubling the performance of benchmark Multi-Criteria Decision Analysis methods (39.3%).
Complementing this spatial analysis, we developed a temporal, area-based melt model from annual inventories of over 2,000 individual glacier polygons. Using a Huber regression to fit negative power laws to ablation rates, we identified a clear acceleration in retreat for smaller ice bodies, consistent with albedo-ice feedback mechanisms. Between 2016 and 2023, we observed a relative area loss of 15 ± 4% (180 ± 70 km2).
Integrating these models to forecast future scenarios, we project that only ~30% (26–43%) of the 2020 glacial surface area will remain by 2100, with several cordilleras facing near-total extinction. This workflow establishes a new standard for observation-based, scalable glacial modelling, providing the high-resolution spatial and temporal insights necessary for effective water resource management and adaptation strategies in the tropical Andes.
How to cite: Lepage, H., Andriychenko Leonenko, D., Elliott, N., and Barnes, C.: Machine Learning and Remote Sensing Projections for Peruvian Glaciers (2016–2100), EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-14578, https://doi.org/10.5194/egusphere-egu26-14578, 2026.
Modelling water infiltration in unsaturated soils is vital for maintaining a healthy ecosystem, analysing the stability of slopes, or promoting sustainable agriculture. Recently, Physics-informed Neural Networks (PINNs) have gained popularity in solving highly nonlinear problems like the Richardson-Richards equation (RRE), by approximating physical laws with a loss term in a mesh-free approach, often using sparse data points, to mimic the gap spacing between field sensors. However, despite several successful applications in modelling 1D infiltration problems, the generalisation capability of these models is often limited by the specific scenarios used during training. Therefore, potential of the neural networks as universal approximators are not exploited in such applications. This paper investigates the feasibility of applying a Parameterised-PINNs (P-PINNs) as a surrogate model to solve the RRE. The model was trained only once across a range of infiltration conditions defined by varying soil hydraulic properties and meteorological conditions to evaluate its ability to predict various scenarios within the multidimensional parameter space without additional observation data. Results show that a wider rather than a deeper network architecture, enhanced by dynamic adaptive techniques, such as time-stratified Residual-based Adaptive Refinement (RAR), Layer-wise Locally Adaptive Activation Function (L-LAAF), and Principled Loss Function (PLF), aids in capturing the correct physical profile. Although the model achieved high overall performance when validated against analytical solutions, Nash-Sutcliffe Efficiency (NSE) > 0.99, it exhibited very minor phase errors. P-PINN was tested across drastically changing parameters, e.g. soils with very high or very low air-entry values, and satisfactory validation metrics were obtained. Our implementation P-PINNs demonstrate the potential as a universal non-linear approximator for such problems, where the initial computational cost of training is offset by the instant large-scale evaluations.
How to cite: Gowely, M. and Yildiz, A.: Parameterised PINNs for water infiltration in unsaturated soils, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8225, https://doi.org/10.5194/egusphere-egu26-8225, 2026.
High-resolution and temporally consistent satellite observations are essential for effectively monitoring, modeling, and mitigating environmental challenges. However, optically based remote sensing faces cross-sensor interoperability issues and is inherently affected by cloud contamination and atmospheric interference, resulting in temporal discontinuities that limit the availability of timely and uninterrupted observations. Existing approaches have primarily focused on retrospective gap-filling of missing data. In contrast, forecasting surface dynamics introduces additional challenges, particularly the need for high-fidelity and temporally continuous information to support near-real-time monitoring and predictive applications as the time since the last observation increases.
To address this challenge, we developed a physics-guided transformer framework trained on Harmonized Landsat, Sentinel-2, and PlanetScope (HLSP) data to forecast uninterrupted daily 30-m surface reflectance during periods with missing optical observations. HLSP is a radiometrically and geometrically harmonized multi-sensor optical dataset integrating Landsat 8–9, Sentinel-2, and PlanetScope imagery to provide sensor-agnostic, temporally consistent surface reflectance products. The model was trained using a multi-year (2017–2025) archive of HLSP surface reflectance imagery across eight agricultural regions in the United States, Brazil, France, Spain, Egypt, South Africa, Thailand, and China. Spectral features from daily HLSP data (30 m resolution) were combined with daily land surface temperature (LST) and soil water content (SWC) at 100-m resolution derived from passive microwave observations. Additional temporal covariates, including day-of-year encoded using sine and cosine transformations, were incorporated to explicitly represent seasonal and phenological timing and enable the network to capture key biophysical, hydroclimatic, and seasonal controls on surface reflectance dynamics.
The physics-guided framework constrains predictions using land–surface energy balance relationships linking surface reflectance, land surface temperature, and soil moisture. These constraints promote physically consistent interactions among surface variables while learning temporally coherent surface reflectance dynamics associated with vegetation growth, moisture persistence, and land–surface energy exchanges.
Model skill was evaluated using RMSE and MAE under a forward-looking temporal validation strategy, in which the model was trained on eight years of historical HLSP data and used to forecast surface reflectance over multiple lead times (2, 5, 10, 15, and 20 days) following the last available optical observation in the final year. Forecasts were validated against independently observed HLSP data for the corresponding periods, allowing assessment of skill degradation as forecast horizons increased. Results demonstrate that incorporating LST, SWC, NDVI, and time-related covariates substantially improves forecast stability and fidelity, particularly under variable climatic and land-cover conditions. The proposed approach provides a scalable and generalizable machine-learning framework for short-term forecasting of EO surface reflectance time series, with applications in climate-impact assessment, drought monitoring, evapotranspiration modeling, and carbon–water flux analysis.
How to cite: Alamdar, S. and Houborg, R.: Physics-Guided Transformer-based Forecasting of High-Resolution Earth Observation Surface Reflectance Data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-14919, https://doi.org/10.5194/egusphere-egu26-14919, 2026.
Accurate monitoring of ground-level air pollutants is essential for exposure assessment and air quality management, but conventional modeling approaches exhibit significant limitations. Chemical Transport Models (CTMs) are computationally intensive and prone to systematic bias, while data-driven models often lack physical consistency and poorly represent long-range transport. To address these limitations, we present a novel hybrid modeling framework with three key innovations. First, satellite retrievals are employed as primary predictors rather than CTM outputs, thereby reducing computational demands. Second, a dual-target learning strategy prioritizes satellite-to-surface relationships, while CTM outputs are incorporated as soft physical constraints in data-sparse regions. Third, a generative diffusion model is integrated to improve the representation of long-range pollutant transport. Focusing on nitrogen dioxide (NO2), the completed framework achieves superior daily predictive accuracy (R2 = 0.72, RMSE = 3.70 ppb), outperforming precursor models. Its successful extension to sulfur dioxide (SO2) and fine particulate matter (PM2.5) demonstrates broad applicability. This study provides a physically informed and computationally efficient solution for scalable generation of high-fidelity, spatially continuous ground-level air quality fields.
This work was funded by the Korea Meteorological Administration Research and Development Program under Grant RS-2024-00404042 and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS-2024-00343921).
How to cite: Kim, H., Kim, E., Kang, Y.-H., Jeong, S., Kim, S., Kim, H. C., and Son, R.: CTM-Assisted Generative AI Framework for Satellite-to-Surface Estimation of Ground-Level Air Pollutants, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8476, https://doi.org/10.5194/egusphere-egu26-8476, 2026.
Machine learning models have been used with success to produce accurate river discharge forecasts at multiple lead times. However, almost no research has been done to show if they are physically consistent across lead times. In the deterministic problem setting, where models output a single forecast with multiple leadtimes, these models are known to be mean-seeking, predicting the most likely river flow for each day, regardless of how likely the resulting trajectory is to occur. This is important for forecasters who need to look at the multi-day properties of a forecast, such as the accumulated flow or number of days over threshold. When each leadtime is described as an independent distribution, the model provides no insight into how to connect the uncertainties at each lead time, as an ensemble forecast would. In this paper, we show that temporal consistency in machine learning forecasts cannot be assumed, and develop two methods for enforcing temporal consistency, the Conditional-LSTM and Seeded-LSTM. Through this, we create ensemble forecasts that successfully predict temporal properties of the 10-day hydrographs. We find that by explicitly training the model to treat the prediction of previous lead times as truth, our model better predicts temporal properties of 10-day hydrographs than other standard methods. Our approach allows users to efficiently generate as many ensemble members as desired, and we use our results to highlight the important of developing temporally consistent ensembles.
How to cite: Ruparell, K., Hunt, K., Cloke, H., Prudhomme, C., Pappenberger, F., and Chantry, M.: AI-generated ensemble river flow forecasting: Using rollout and an additional noise input to build ensemble forecasts, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2804, https://doi.org/10.5194/egusphere-egu26-2804, 2026.
The availability of fresh water is vital to the ecosystem and communities. In a changing climate, the increased risk of droughts makes it crucial to have an accurate understanding of changes in terrestrial water storage (TWS). Predicting changes in TWS is inherently difficult since it integrates the changes of all water compartments, with underlying processes that operate on vastly different temporal and spatial scales.
Forecasting tasks nowadays are often solved using machine learning models. However, these models require vast amounts of data. In contrast, total water storage anomalies (TWSA) derived from GRACE/GRACE-FO observations only date back to 2002 and are available at a grid of 1°x1° at monthly resolution. Nevertheless, Li et al., (2024) showed that machine-learning approaches could forecast TWSA tendencies for up to one year ahead. They cleverly exploit temporal lag relationships between TWSA and ocean, atmospheric, or land variables.
In our work, we explore a novel design of a hierarchical graph using domain knowledge of hydrological basins to encode these processes in a latent feature sequence using an encoder-processor-decoder style graph neural network. The subsequent recurrent neural network then forecasts TWSA from the latent feature sequence and 12-month history of TWSA for up to six months ahead. The gridded product of the seasonal forecast of global TWSA shows improvement over a seasonal long-term mean.
Li, F., Kusche, J., Sneeuw, N., Siebert, S., Gerdener, H., Wang, Z., Chao, N., Chen, G., and Tian, K.: Forecasting Next Year’s Global Land Water Storage Using GRACE Data, Geophys. Res. Lett., 51, e2024GL109101, https://doi.org/10.1029/2024GL109101, 2024.
How to cite: Steidl, V. and Zhu, X. X.: Hierarchical Graph Networks for ForecastingTerrestrial Water Storage Anomalies, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18659, https://doi.org/10.5194/egusphere-egu26-18659, 2026.
Mapping the Peruvian Andes has high ecological value because its ecosystems are immensely diverse. These mountains shelter numerous endemic species that could be protected if informed decisions are made when delineating conservation zones. Rigorous analysis of high-altitude regions traditionally requires multiple field visits, which place a financial burden on research teams. Such visits can pose safety risks, as several remote areas are difficult to access on foot due to the steep gradients, cloud cover, and logistical limitations.
Recent advances in satellite missions and machine learning (ML) allow land-cover features to be characterised with fewer ground-truthing expeditions, by utilising patterns present in large imagery datasets. However, the Andes remain challenging to map, because of the spectral similarity among some land-use and land-cover (LULC) classes and because steep gradients can lead to geometric distortions in the recorded images.
This study highlights an easy-to-use method for generating LULC map prototypes for high-altitude Andean regions using EnMAP and EMIT hyperspectral imagery (HSI). Machine learning algorithms (e.g., K-means clustering, principal component analysis) were applied to the HSI to generate clusters and extract features with high discriminant power among LULC types. Expert interpretation allowed pairing the obtained clusters with suitable ecosystem labels, producing prototype LULC maps.
How to cite: Radu, D.-I., Lepage, H., Barnes, E., and Barnes, C.: Mapping Ecosystems in the Peruvian Andes Using Hyperspectral Imagery and Machine Learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9777, https://doi.org/10.5194/egusphere-egu26-9777, 2026.
Accurate storm surge prediction is essential for reducing the risks associated with extreme sea levels and for supporting early warning and preventive measures. Physically based numerical models continue to improve in skill and resolution, but their high computational cost limits their use in large ensembles and long-term scenario analyses. Recent advances in machine learning offer a complementary pathway for efficient storm surge forecasting. Here, a machine-learning framework is developed, calibrated, and validated to predict extreme sea levels in the North Sea and Baltic Sea. The model is based on 58 years of spatially distributed wind data and uses a Long Short-Term Memory (LSTM) architecture to capture the temporal dynamics driving water level variability. Compared to traditional physically based hydrodynamic models, the machine-learning approach requires only a fraction of the computational resources, enabling rapid probabilistic and large-ensemble forecasts across large domains and extended time periods. This efficiency is particularly valuable for climate change research, where large ensembles are generally needed to address the combined uncertainty of climate and hydrodynamic models but remain computationally prohibitive using conventional approaches. By providing a scalable and resource-efficient alternative, this framework enables consistent storm surge prediction across timescales ranging from short-term forecasting to long-term climate projections over decades.
How to cite: Mik-Meyer, V., Pereira, F. C., Larsen, M. A. D., Su, J., and Drews, M.: Seamless Storm Surge Prediction Using a Surrogate Hydrodynamic Model Based on Long Short-Term Memory Networks, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-22145, https://doi.org/10.5194/egusphere-egu26-22145, 2026.
Orals: Fri, 8 May, 08:30–09:40 | Room -2.15
Machine learning hold the promise of unlocking more accurate and realistic parameterizations of atmospheric processes, but brings its own set of challenges and drawbacks. Among top issues are generalization, stability and interpretability. Here we present a parameter-efficient neural network parameterization which aims to address these issues by incorporating physical knowledge to a high degree. By predicting fluxes and microphysical process rates instead of total tendencies, the conservation of water can be hardcoded, which is shown to improve online performance. Furthermore, a physically motivated architecture based on vertically recurrent neural networks enables high computational efficiency and a low number of parameters. The models are trained and evaluated using a superparameterization setup with real orography. The impact of incorporating stochasticity is also discussed.
How to cite: Ukkonen, P. and Christensen, H.: Physics-informed, open-box neural network parameterization of moist physics, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13711, https://doi.org/10.5194/egusphere-egu26-13711, 2026.
As the spatial resolution of general circulation models (GCMs) increases and storms and clouds can be resolved, the underlying cloud microphysics still need to be parameterised. This is known to be a major source of uncertainty in climate and weather simulations. The established parameterisations use bulk moment schemes, where the conversion of cloud and rain droplets is approximated through empirical relationships. Particle-based superdroplet simulations would provide a more accurate representation but are typically not feasible for use in GCMs.
We couple SuperdropNet, an ML emulator for warm rain cloud microphysics trained on superdroplet simulations, to ICON. Previously, we validated the coupled model in an idealised cloud microphysics test case and showed that SuperdropNet runs stable and provides reasonable precipitation patterns.
Now we move towards a climate model experiment with 10 km horizontal resolution in an AMIP setup to investigate SuperdropNet’s feasibility and interaction with ICON in a realistic setting. Coupling SuperdropNet to ICON is achieved using FTorch. We are able to run ICON on 128 nodes on the CPU partition of the HPC system Levante with minimal overhead. Conditions beyond the training data range of SuperdropNet lead to negative feedback loops and impact the long-term stability of the coupled simulation. Therefore, we implement physics-based constraints that improve stability. Initial results show mean surface precipitation is very similar to using the bulk scheme approach. SuperdropNet simulates a faster cloud-to-rain transition which impacts cloud water mass and rain droplet size. This has consequences for the radiation budget and the frequency distribution of precipitation. Furthermore, we show that an autoregressive rollout of SuperdropNet that allows for longer GCM time steps runs stable and does not impact results. Finally, we test SuperdropNET’s generalisation capabilities in a 4K warmer world.
How to cite: Keil, P., Arnold, C., and Sharma, S.: Coupling ICON with a Machine Learning Emulator for Cloud Microphsysics, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18204, https://doi.org/10.5194/egusphere-egu26-18204, 2026.
The use of machine learning to represent small-scale processes, such as subgrid-scale (SGS) dynamics, is now well established in weather forecasting and climate modelling. Recent advances have demonstrated that SGS models trained via "online" end-to-end learning - where the dynamical solver operating on the filtered equations participates in the training - can outperform traditional physics-based approaches. However, most studies have focused on idealised periodic domains or spheres, neglecting mechanical boundaries present in systems such as planetary interiors. To address this issue, we introduce a pseudo-spectral differentiable solver for the study of two-dimensional quasi-geostrophic turbulence in a rapidly rotating, axially symmetric bounded domain. A key advantage of the online learning approach is its implicit correction of the commutation errors arising from the irregular Chebyshev grid used in the radial direction, achieved through the estimation of correction terms for the filtered equations. In addition, since Chebyshev polynomials are not boundary-preserving, we project training data extracted from the high-resolution direct numerical simulation (DNS) from the fine grid onto the coarse grid using a Galerkin approach that ensures compatibility with the boundary conditions.
We examine three configurations, varying the geometry (between an exponential container and a spherical shell) and the rotation rate. The flow is driven by a prescribed analytical forcing that mimics a network of pumps, allowing precise control over the energy injection scale and an exact estimate of the power input. For each case, we evaluate the accuracy of the online-trained SGS model against the reference DNS using integral quantities and spectral diagnostics. In all configurations, we show that an SGS model trained on data spanning only one turnover time remains stable and accurate over integrations at least a hundred times longer than the training period. Moreover, we demonstrate the model's remarkable ability to reproduce slow processes occurring on time scales far exceeding the training duration, such as the inward drift of jets in the spherical shell geometry, which exhibits a quasi-periodic recurrence time of O(10) turnover times. These results suggest a promising path towards developing SGS models for planetary and stellar interior dynamics, including dynamo processes. They indicate that costly DNS may need to be run only for short durations to generate training data, enabling subsequent long-term simulations with the trained model at a negligible computational cost.
How to cite: Fournier, A., Frezat, H., and Gastine, T.: Online learning of subgrid-scale models for quasi-geostrophic turbulence in planetary interiors, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4279, https://doi.org/10.5194/egusphere-egu26-4279, 2026.
Estimating the source of acoustic waves propagating in a vertically stratified medium poses significant challenges due to the high variability of the acoustic fields at long-range distances caused by heterogeneous vertical sound speed profiles. This renders the problem an inverse and ill-posed one. To address this challenge, we propose a three-step approach utilizing a Bayesian framework. First, we show that using only the low-frequency components (up to 1.5 Hz) of the acoustic fields is sufficient to capture the source parameters. Second, we develop a fast surrogate forward model based on a Fourier Neural Operator (FNO) [1] to bypass the computational burden of traditional numerical solvers. Finally, we trained diffusion models to represent the complex prior [2] of the atmospheric profiles and to accurately estimate the posterior distribution [3] in the context of our inference problem. The models are trained on a database comprising over 20,000 simulations generated using a normal mode code [4]. Our results show that our FNO model achieves a relative least squares error of approximately 8%. The combined FNO and diffusion model framework [5] is demonstrated to yield more reliable energy estimates when compared to the utilization of the FNO framework alone.
[1] N. Perrone, F. Lehmann, H. Gabrielidis, S. Fresca, and F. Gatti, “Integrating Fourier Neural Operators with Diffusion Models to improve Spectral Representation of Synthetic Earthquake Ground Motion Response,” arXiv preprint arXiv:2504.00757, 2025. doi: 10.48550/arXiv.2504.00757
[2] F. Lehmann, F. Gatti, M. Bertin, and D. Clouteau, “3D elastic wave propagation with a Factorized Fourier Neural Operator (F-FNO),” Computer Methods in Applied Mechanics and Engineering, vol. 417, art. no. 116718, 2023. doi: 10.1016/j.cma.2023.116718
[3] F. Bergamin, C. Diaconu, A. Shysheya, P. Perdikaris, J. M. Hernández-Lobato, R. E. Turner, and E. Mathieu, “Guided Autoregressive Diffusion Models with Applications to PDE Simulation,” in ICLR 2024 Workshop on AI4DifferentialEquations In Science, 2024.
[4] T. Karras, M. Aittala, S. Laine, and T. Aila, “Elucidating the design space of diffusion-based generative models,” in Proceedings of the 36th International Conference on Neural Information Processing Systems (NeurIPS), 2022, pp. 26565–26577. doi: 10.5555/3600270.3602196
[5] M. Bertin, C. Millet, and D. Bouche, “A low-order reduced model for the long range propagation of infrasound in the atmosphere,” The Journal of the Acoustical Society of America, vol. 136, no. 5, pp. 2693–2705, 2014. doi: 10.1121/1.4896776
How to cite: Noëlé, E., Gatti, F., Clouteau, D., Millet, C., and Lehmann, F.: A Hybrid FNO-Diffusion Framework for Uncertainty-Aware Source Energy Estimation in Atmospheric Waveguides, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-14900, https://doi.org/10.5194/egusphere-egu26-14900, 2026.
How to cite: Wang, X., Mengaldo, G., Chen, J., Yang, J., Adie, J., See, S., Furtado, K., Chen, C., Arcomano, T., Maulik, R., and Xue, W.: CondensNet: Self-adaptive physical constraints for stable long-term hybrid climate simulations, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16098, https://doi.org/10.5194/egusphere-egu26-16098, 2026.
The primary challenge in forecasting the Earth's magnetic field lies in capturing rapid, non-linear events like geomagnetic jerks. Conventional models relying solely on geomagnetic data often struggle to replicate the variation of those abrupt changes, such as the 2014–2015 geomagnetic jerk.
This study introduces a multiple data approach that simultaneously co-estimates geomagnetic snapshots and Length-of-Day (LOD) variations using a machine learning method. Specifically, we use an Extended Kalman Filter-trained Recurrent Neural Network (EKF-RNN; Sato et al., in press) to model the complex, non-linear dynamics of the Earth's core, including the geomagnetic jerks.
The training and validation datasets for our neural network were derived from the MCM geomagnetic field model (Ropp & Lesur, 2023), which is based on vector geomagnetic data from global magnetic observatories as well as the CHAMP and Swarm-A satellites (Ropp et al., 2020). To constrain the internal dynamics of the Earth’s core, we incorporated LOD data from the Earth Orientation Parameters series C04, provided by the International Earth Rotation and Reference Systems Service. The LOD dataset combines historical observations with modern space geodetic techniques including Very Long Baseline Interferometry, Satellite Laser Ranging, Global Navigation Satellite Systems and Lunar Laser Ranging, offering a continuous record from 1962 to present (Bizouard & Gambis, 2011).
After removing predictable tidal and atmospheric signals, LOD variations reflect exchanges of angular momentum between the Earth's core and mantle. Since electromagnetic waves such as torsional Alfvén waves generated in the Earth's core are linked to rapid geomagnetic accelerations, inclusion of LOD data may make a key constraint on the geomagnetic forecast. Our results show that a model trained only by geomagnetic secular acceleration (SA) failed to capture the 2014–2015 geomagnetic jerk, whereas adding LOD data showed an improved accuracy during the same event. Specifically, the SA misfit decreased from 4.98 to 2.43 nT/yr². The improvement was most significant when training with the second-order derivatives (i.e., SA snapshots themselves), indicating that the EKF-RNN successfully uncovered the underlying physical connection between geomagnetic acceleration and the Earth’s rotation. This study confirms that a multiple data approach, combining independent yet physically linked observation data, is essential for the next generation of geomagnetic forecast models.
How to cite: Toh, H., Sato, S., and Lesur, V.: Impact of Length-of-Day inclusion on geomagnetic secular variationforecast by a recurrent neural network, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15244, https://doi.org/10.5194/egusphere-egu26-15244, 2026.
Canopy resistance is a key parameter in the Penman–Monteith (P–M) equation for calculating evapotranspiration (ET). In this study, we compared a machine learning algorithm–support vector machine (SVM) and an analytical solution (Todorovic, 1999) for estimating canopy resistances. Then, these estimated canopy resistances were applied to the P–M equation for estimating ET; as a benchmark, a constant (fixed) canopy resistance was also adopted for ET estimations. ET data were measured using the eddy-covariance method above three sites: a grassland (south Ireland), Cypress forest (north Taiwan), and Cryptomeria forest (central Taiwan) were used to test the accuracy of the above two methods. The observed canopy resistance was derived from rearranging the P–M equation. From the measurements, the average canopy resistances for the grassland, Cypress forest, and Cryptomeria forest were 163, 346, and 321 (s/m), respectively. Our results show that both methods tend to reproduce canopy resistances within a certain range of intervals. In general, the SVM model performs better, and the analytical solution systematically underestimates the canopy resistances and leads to an overestimation of evapotranspiration. It is found that the analytical solution is only suitable for low canopy resistance (less than 100 s/m) conditions.
How to cite: Hsieh, C.-I., Huang, I.-H., and Lu, C.-T.: Estimating Canopy Resistance Using Machine Learning and Analytical Approaches, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2240, https://doi.org/10.5194/egusphere-egu26-2240, 2026.
Forecasting the location and intensity of strong winds associated with midlatitude cyclones remains a key challenge due to their substantial societal and environmental impacts. In this study, we conditionally evaluate the ability of numerical weather prediction (NWP) models and machine learning weather prediction (MLWP) models to represent wind structures linked to these cyclones. Using a feature‑based tracking approach applied to a large sample of Northern Hemisphere cyclone events, we compare how different modelling frameworks capture cyclone evolution, including track, intensity, and near‑surface wind characteristics.
Our analysis shows that MLWP models can reproduce broad aspects of cyclone behaviour, such as large‑scale track evolution, with skill comparable to established operational NWP forecasting systems at medium-range lead times. However, we also identify systematic differences in how these models represent cyclone intensity and associated wind extremes. In particular, MLWP models tend to underestimate key high‑impact features, such as minimum pressure and peak near‑surface winds, relative to dynamical NWP forecasts.
These findings highlight both the promise and current limitations of MLWP systems for predicting midlatitude cyclone hazards. Understanding these behaviours provides guidance for future model development and for the use of ML‑based forecasts in operational and risk‑focused applications.
How to cite: Dacre, H., Charlton-Perez, A., Driscoll, S., Gray, S., Harvey, B., Harvey, N., Hodges, K., Hunt, K., and Volonte, A.: Midlatitude Cyclone Intensity Biases in Machine Learning Weather Prediction Models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3361, https://doi.org/10.5194/egusphere-egu26-3361, 2026.
The inherent limitations of individual geophysical methods and the sparsity of observational data often render inversion results unstable and non-unique. Joint inversion of multiphysics data exploits the complementary sensitivities of different physical fields regarding depth, resolution, and boundary features, thereby significantly mitigating the ambiguity of single-method inversion and enhancing interpretation reliability. Traditional joint inversion approaches primarily fall into two categories: spatial structure-based and physical parameter-based constraints. The former relies on the similarity of property distribution patterns, which struggles to decouple non-homologous anomalies, while the latter is often constrained by the unreliability of empirical relationships under complex geological conditions. Recently, deep learning methods based on the U-Net architecture have achieved joint inversion by establishing constraints based solely on spatial structural similarity (Hu et al., 2025) or physical parameter correlations (Guo et al., 2021). Although promising, these methods often fail to accurately characterize non-homologous anomalies in complex geological environments.
This study proposes a dual-stream 3D U-Net architecture incorporating a hybrid attention-gating mechanism. In terms of methodology, we first construct a training dataset based on rock physics data that encompasses both statistical correlations and structural discrepancies. Regarding the network architecture, independent encoders are employed to extract 3D features from gravity and magnetic data, respectively. A cross-attention module is then utilized to capture deep structural correlations, thereby enhancing cooperative inversion in homologous regions. Subsequently, a gated fusion module is introduced as an adaptive feature selector to effectively disentangle inconsistent features in non-homologous regions. Finally, the prediction models are generated through independent decoders.
During the joint inversion implementation phase, the network takes preliminary independent inversion results as input to predict high-fidelity models that integrate physical and geological priors. We incorporate these predicted models as reference models into the regularization term of the joint inversion objective function, constructing a deep-prior-based constraint. During iterative optimization, this constraint guides the inversion trajectory toward the fine geological structures predicted by deep learning by minimizing the discrepancy between the inverted and reference models, while ensuring the fit to observational data. This mechanism achieves an organic integration of data-driven and physics-driven approaches.
References
- Hu, Y. Su, X. Wu, Y. Huang and J. Chen, "Successive Deep Perceptual Constraints for Multiphysics Joint Inversion," in IEEE Transactions on Geoscience and Remote Sensing, vol. 63, pp. 1-14, 2025, Art no. 5907114.
- Guo, H. M. Yao, M. Li, M. K. P. Ng, L. Jiang and A. Abubakar, "Joint Inversion of Audio-Magnetotelluric and Seismic Travel Time Data With Deep Learning Constraint," in IEEE Transactions on Geoscience and Remote Sensing, vol. 59, no. 9, pp. 7982-7995, Sept. 2021.
How to cite: Xi, B. and wang, Z.: A Hybrid Attention-Gating Deep Learning Framework for 3D Joint Inversion of Gravity and Magnetic Data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5241, https://doi.org/10.5194/egusphere-egu26-5241, 2026.
The radiation belts are populations of energetic particles, such as electrons and protons, trapped in the near-Earth space vicinity by the geomagnetic field. Because they cover the great majority of existing orbits and because the particles’ dynamics, highly coupled with solar activity, can strongly affect spacecraft components and mission, the accurate modeling of these regions is of uttermost importance for the monitoring of the near-Earth space dynamics.
Traditionally, the radiation belts are modeled by solving a three‑dimensional diffusion equation with numerical solvers. While a single 3D simulation can easily be run in real time, as done routinely in many space weather forecasting pipelines, the computational burden can become significant when the model is used in ensemble‑based data assimilation that potentially requires hundreds of runs, over very long periods, such as those needed for space‑climate studies.
Within this context, machine learning based Reduced Order models (ROMs) offer an interesting solution to approach the solutions of traditional high-fidelity physics-based models with a reasonable accuracy and at a reduced computational cost. This is achieved by projecting project highly non-linear features onto a disentangled, interpretable latent space of reduced dimension which dynamics could be driven by external variables.
In this work, we take a first step towards the development of a ROM for the Earth electron radiation belts. Using a Distance regularized Siamese twin autoencoder (DIRESA) on long-term simulations we manage to reduce electron fluxes on a refined grid to a small subset of latent variables. We then show that these variables that can all be linked with external geomagnetic parameters. This allows them to be at the core of a ROM of the Earth electron radiation belts driven by those external parameters.
This work was supported by both the "Event-Based Electron Belt Radiation Storm Environments Modelling" Activity led by the Space Applications & Research Consultancy (SPARC) under ESA Contract 4000141351/23/UK/EG and ONERA internal fundings, through the federated research project PRF-FIRSTS.
How to cite: Nguyen, G., Brunet, A., Tahtouh, M., Bernoux, G., Dahmen, N., and Sandberg, I.: Disentangled and interpretable feature extraction of the Earth electron radiation belt: a first step towards the development of a reduced order model, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11323, https://doi.org/10.5194/egusphere-egu26-11323, 2026.
Reduced-order surrogate models based on Koopman autoencoders have recently shown strong potential for accelerating flexible-mesh coastal ocean simulations while maintaining physically meaningful dynamics. In this contribution, we extend a previously validated Koopman autoencoder framework by explicitly incorporating information from in-situ measurements during training.
The proposed approach augments the surrogate training objective with measurement-based constraints, penalizing deviations from observed water surface elevations at selected locations and times. This enables the surrogate to remain consistent with sparse observations while preserving the learned large-scale dynamical structure driven by meteorological forcing and boundary conditions.
The method is evaluated on two realistic MIKE 21 HD coastal-ocean configurations published as open WaterBench datasets: the Southern North Sea and the Øresund Strait. Performance is assessed against both full physics-based simulations and independent in-situ observations, focusing on accuracy, temporal stability, and generalization beyond the training period.
Results demonstrate that measurement-constrained training can reduce local prediction errors near observation points without degrading global performance, while retaining the substantial inference speed-ups characteristic of Koopman-based reduced-order models. The proposed framework represents a step toward tighter integration of observations and machine-learning surrogates for efficient, observation-aware coastal ocean modelling, with relevance for ensemble forecasting and long-term scenario analysis.
How to cite: Jessen, M. U., Mariegaard, J. S., and Petersen, F. H.: Measurement-Constrained Reduced-Order Surrogates for Flexible-Mesh Coastal Ocean Models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11847, https://doi.org/10.5194/egusphere-egu26-11847, 2026.
We demonstrate software and tools for users to progress from machine learning theory, probabilistic methods, through to construction of AI models across environmental science. We span basic AI methods through to modern generative AI methods, physics informed techniques, as well as including a vast array of concrete applications such as river discharge modelling, ocean-wave emulation, environmental monitoring, AI foreasting and more. Throughout we place emphasis on how and when these methods should be used, as well as their limitations. This allows users to develop a non-naive understanding of AI and to engage with all themes of Machine Learning Across Earth System Modeling: Subgrid-Scale Parameterizations, Emulation and Hybrid Modeling.
How to cite: Driscoll, S., Hunt, K., Mansfield, L., Swaminathan, R., Wei, H., Bach, E., and Peard, A.: Weather and Climate: Applications of Machine Learning and Artificial Intelligence, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21143, https://doi.org/10.5194/egusphere-egu26-21143, 2026.
While numerical weather prediction (NWP) underpins existing early warning systems, its high computational cost limits scalability. Machine-learning weather prediction (MLWP) offers a promising alternative, yet its skill and reliability at forecasting wind extremes and small-scale storm features across different storms remain uncertain. Evaluating the forecast skill of MLWP models across a range of storms is therefore critical before MLWP can be integrated safely into early warning systems.
This study evaluates the performance of eight leading MLWP models at forecasting the peak 10 m and 850 hPa wind speeds, pressure minima, and relative vorticity associated with the most damaging UK windstorms from the 2023/24 winter season: Babet, Ciarán, Debi, Gerrit, Henk and Isha. MLWP models are evaluated against ERA5 and IFS analysis and benchmarked against the NWP IFS ensemble forecast. The results reveal substantial variability in MLWP forecasting skill both between storms and across models.
MLWP forecast skill is found to be linked to the horizontal scale and dynamical nature of the storm feature producing the strongest winds. While wind maxima associated with large-scale conveyor-belt airstreams are generally well predicted, those arising from smaller-scale features, including the cold conveyor belt and sting jets, are underestimated. MLWP model performance is also found to be variable between storms, with no clear best- or worst-performing model. The higher-resolution Aurora-0.1 model is not found to be better at forecasting wind extremes, despite the small spatial scale of the storm features producing the strongest winds in four of the storms analysed.
An in-depth, feature-based analysis is performed for Storms Henk and Isha. Henk proved challenging for both MLWP and NWP models to forecast, resulting in short-notice and inaccurate wind alerts from the Met Office. The MLWP models performed worst for Isha overall, despite the NWP models predicting it well. Across both storms, MLWP models struggled to predict small-scale features associated with extreme winds and tended to smooth sharp frontal gradients.
These results highlight critical limitations in existing MLWP models that make them unsuitable for replacing NWP as a primary forecasting tool for hazardous UK windstorms today. However, current MLWP models could provide rapid, low-cost ensemble information that complements traditional NWP outputs, or serve as a part of a hybrid ML-NWP approach, particularly if structural limitations in representing fine-scale wind maxima are acknowledged and mitigated.
How to cite: Hewitt, J., Volonté, A., Harvey, B., Andrade Cardoso, A., Hunt, K., Harvey, N., Martinez-Alvarado, O., Gray, S., Dacre, H., and Hodges, K.: Evaluating the forecast skill of machine-learning weather prediction models across a selection of extreme UK windstorms, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12760, https://doi.org/10.5194/egusphere-egu26-12760, 2026.
Machine-learning emulators have demonstrated remarkable skill for weather prediction and short-range forecasting, yet their behaviour, as forecasts extend toward seasonal and longer timescales, remains less well explored and understood. Approaching these horizons, forecast skill is shaped less by short-range error growth, while variations in background states or system parameters increasingly influence the evolving dynamics. Understanding if and how different neural architectures perform with such changes is therefore central to assessing their suitability for emulation beyond medium range weather prediction, where robustness plays an increasingly important role. In this work, we investigate how inductive biases encoded in deep-learning architectures influence their ability to represent and evolve dynamics as forecasts move into windows nearing and sometimes beyond their training data.
We use the idealised climate model ISCA as a controlled testbed, enabling systematic variation of planetary parameters and initial conditions while retaining a fixed underlying set of governing equations. Emulators are trained on ensembles of trajectories sampled from a restricted parameter range and evaluated under progressively more challenging ID/OOD settings. This framework allows us to disentangle errors arising from finite-horizon forecasting from those associated with longer-timescale dynamical shifts, providing insight into which architectural biases promote stability, physical consistency, and robustness as machine-learning models are pushed from shorter term prediction toward longer time scale emulations.
How to cite: Bohn Lassen, O., Camara Pereira, F., Driscoll, S., Schemm, S., and Thomson, S.: Inductive Biases for Robust Climate Emulation Across Forecast Timescales, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16240, https://doi.org/10.5194/egusphere-egu26-16240, 2026.
AI weather models and weather-based foundation models have demonstrated impressive skills in short- to medium-range forecasts. While most weather models become unstable on longer time scales, a wide variety of AI climate emulators have been proposed, raising questions about the fundamental differences between these approaches.
In this work, we compare state-of-the-art models when producing rollouts on annual time scales. We quantify and characterize different types of instability: smoothing, visual artifacts, drift, and loss of seasonality. This analysis highlights the previously unreported stability of the Aurora foundation model and the Earth System Foundation Model (ESFM) for rollouts longer than 35 years.
To encompass more diverse representations of possible states of the Earth, ESFM is pretrained on a variety of CMIP6 datasets from the historical period, in addition to the ERA5 reanalysis commonly used in AI models. ESFM also includes climate forcings for physically driven long rollouts. We demonstrate the benefits of CMIP6 pretraining when finetuning on new CMIP6 datasets, including datasets with higher resolution, unseen physical processes, and climate change scenarios.
Overall, this work opens perspectives to adapt large-scale pretrained foundation models to the specific challenges of climate projections.
How to cite: Lehmann, F., Neumarker, R., Scorrano, G., Cheng, Y., Mohebi, S., Ozdemir, F., Gou, J., Fuhrer, O., Hoefler, T., Mishra, S., Salzmann, M., Schemm, S., and Soja, B.: Extending foundation models from weather to climate: challenges and promises, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19872, https://doi.org/10.5194/egusphere-egu26-19872, 2026.
Shadows cast by terrain and tall structures are a persistent limitation in satellite imagery, since they degrade radiometric consistency and compromise downstream tasks such as classification, detection, and 3D reconstruction. In this context, machine learning methods for shadow removal provide a flexible and easy-to-deploy tool to assist satellite remote sensing tasks. Nevertheless, one prominent issue for its development in Earth Observation (EO) is the scarcity of publicly available, geometry-consistent paired shadowed/shadow-free satellite data. Most EO resources support shadow detection or 3D modelling but not shadow removal, while existing shadow-removal datasets largely target ground-level or UAV imagery and do not reflect multi-date, multi-angle satellite acquisition.
To address this gap, we present deSEO, a physics-informed, geometry-aware methodology that converts anyinto paired training data for weakly supervised satellite shadow removal. We exemplified our procedure on the S-EO satellite dataset. Using the multi-temporal, multi-geometry S-EO dataset (WorldView-3 imagery with DSM priors, simulated shadow masks, and RPC camera models), deSEO selects a minimally shadowed acquisition per tile as a proxy reference and pairs it with more shadowed dates under explicit temporal and geometric constraints. Residual off-nadir parallax is mitigated through orientation normalisation and feature-based registration (LoFTR + RANSAC), yielding a per-pixel validity mask that can be used to restrict model supervision to reliably aligned regions.
To validate the usability of the shadow-removal dataset derived from S-EO, we first adapted UAV-oriented methods such as SRNet and pix2pix. However, these approaches fail to converge to a stable training regime under the viewpoint variability typical of satellite acquisitions. We therefore develop a more robust method and training strategy that mitigates this common failure mode of image-to-image translation on multi-date, multi-geometry satellite imagery. Our approach involves training a DSM-conditioned conditional GAN with a U-Net-based generator. The model incorporates perceptual reconstruction and mask-constrained adversarial objectives, with a soft shadow-mask attention prior that emphasises shadow-transition regions. These enhancements overcome the limitations of the classical GAN image translation setup that worked well for UAV data. We evaluate the model on a held-out test split, where the proposed approach achieves a PSNR of 18 ± 1 dB, SSIM of 0.49 ± 0.08, and LPIPS of 0.46 ± 0.05. Notably, improvements were most pronounced at cast-shadow boundaries, and ablation studies revealed that DSM conditioning was the dominant contributing factor, something absent in the SRNet model.
Overall, deSEO provides a reproducible approach to derive paired supervision for satellite shadow removal and establishes a geometry-aware baseline for robust deshadowing under realistic EO acquisition variability.
How to cite: Beltrame, L., Salzinger, J., Fanta-Jende, P., Lampert, J., Thiele, P. L., Svoboda, F., Timofte, R., and Körner, M.: Geometry- and Physics-Aware Dataset Creation for Shadow Removal in High-Resolution Satellite Imagery, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21279, https://doi.org/10.5194/egusphere-egu26-21279, 2026.
As climate change drives more extreme wildfire behavior, accurate and computationally efficient fire spread modeling is increasingly critical for monitoring, mitigation, and risk assessment. Wildfires pose a particularly challenging modeling problem due to their complex interactions with fuels, terrain, and atmospheric conditions, as well as their potential to impact populated regions with severe environmental, economic, and human consequences. These challenges motivate the development of surrogate modeling approaches capable of emulating physics-based wildfire simulations at substantially reduced computational cost. In this work, we present a systematic comparison of two deep learning surrogate model architectures for spatiotemporal wildfire emulation: a convolutional neural network-based generative model and a conditional diffusion model. Both approaches are designed to be grid-invariant and trained to predict three key wildfire variables – time of arrival, flame length, and burn scar – at fixed 15-minute time steps. Model performance is evaluated using an autoregressive rollout procedure in which successive short-term predictions are recursively fed back as inputs to simulate wildfire evolution over 12-hour time horizons. The training data consists of wildfire simulations generated using a Rothermel-based fire spread model with realistic, satellite-derived fuel distributions over the western United States (California and Nevada). Evaluation is performed on geographically distinct fire scenarios to assess generalization across diverse fuel configurations. Both surrogate models are shown to produce stable and physically plausible wildfire dynamics over 12-hour autoregressive rollouts while reducing inference time relative to physics-based solvers. The results highlight the potential of deep generative surrogates to enable rapid ensemble-based risk assessment and support operational fire management workflows under diverse environmental conditions.
How to cite: Boucher, M., Zhang, J., Pain, C., Li, Y., Joshi, A., Moseley, B., and Cunningham, P.: A Comparative Evaluation of Grid-Invariant Deep Learning Surrogate Models for Wildfire Simulation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-22628, https://doi.org/10.5194/egusphere-egu26-22628, 2026.
Improvements of Machine Learning (ML)-based radiation emulators remain constrained by the underlying assumptions to represent horizontal and vertical subgrid-scale cloud distributions, which continue to introduce substantial uncertainties. In this study, we introduce a method to represent the impact of subgrid-scale clouds by applying ML to learn processes from high-resolution model output with a horizontal grid spacing of 5km. In global storm resolving models, clouds begin to be explicitly resolved. Coarse-graining these high-resolution simulations to the resolution of coarser Earth System Models yields radiative heating rates that implicitly include subgrid-scale cloud effects, without assumptions about their horizontal or vertical distributions. We define the cloud radiative impact as the difference between all-sky and clear-sky radiative fluxes, and train the ML component solely on this cloud-induced contribution to heating rates. The clear-sky tendencies remain being computed with a conventional physics-based radiation scheme. This hybrid design enhances generalization, since the machine-learned part addresses only subgrid-scale cloud effects, while the clear-sky component remains responsive to changes in greenhouse gas or aerosol concentrations. Applied to coarse-grained data offline, the ML-enhanced radiation scheme reduces errors by a factor of up to 4-10 compared with a conventional coarse-scale radiation scheme. We observe improved radiative heating rates across several cloud regimes and regions, including precipitating and non-precipitating clouds and stratocumulus regions. This shows the potential of representing subgrid-scale cloud effects in radiation schemes with ML for the next generation of Earth System Models.
How to cite: Hafner, K., Shamekh, S., Bertoli, G., Lauer, A., Pincus, R., Savre, J., and Eyring, V.: Representing Subgrid-Scale Cloud Effects in a Radiation Parameterization using Machine Learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12000, https://doi.org/10.5194/egusphere-egu26-12000, 2026.
Global climate models (GCMs), like KNMI’s EC-Earth, are an important tool to study the global climate system, and to understand how the climate responds to changes in external forcing. Large ensembles of climate simulations are necessary to separate the forced response from fluctuations due to the climate system’s internal variability (Maher et al., 2021; Muntjewerf et al, 2023). GCMs are computationally very expensive to run, particularly as they move towards the km-scale, which makes generating large ensembles very expensive.
The generative modelling framework allows the transformation of a base distribution to the target distribution and easily facilitates the construction of large ensembles. We compare two generative models: 1) “stochastic interpolants”, that learn a pseudo-time dependent stochastic process that directly interpolates between the current state and the conditional target state of interest, and 2) a “flow matching” model, that learns a pseudo-time dependent deterministic process, conditioned on the current state, between a Gaussian distribution and the target state of interest. Both models use a PDE-transformer backbone (Holzschuh et al, 2025).
We train an emulator to predict global 2m-temperature at time t+1 using the previous 5 days of temperature, the annual global mean temperature and some static spatial and temporal features as conditioning inputs. We make predictions auto-regressively, feeding each prediction back into the model to generate sequences of arbitrary length at inference time. We use Large Ensembles from the EC-Earth3 model, for which a transient 16-member (1950-2166) ensemble and two 160-member time slices (2000-2009, 2075-2085) are available (Muntjewerf et al., 2023). The training dataset consists of up to 5 transient members and we use a single member for validation during training. We use another member for inference to produce an ensemble of global temperature simulations.
The flow matching model successfully generates a stable ensemble of temperature fields that simulates the long-term forced trend, interannual variability, and spatial patterns of (global) temperature similarly to the GCM.
References:
Maher, N., Milinski, S. and Ludwig, R., 2021. Large ensemble climate model simulations: introduction, overview, and future prospects for utilising multiple types of large ensemble. Earth System Dynamics, 12(2), pp.401-418.
Muntjewerf, L., Bintanja, R., Reerink, T. and Van Der Wiel, K., 2023. The KNMI Large Ensemble Time Slice (KNMI–LENTIS), Geosci. Model Dev. 16 4581–4597. doi: 10.5194.
Holzschuh, B., Liu, Q., Kohl, G., & Thuerey, N. (2025). PDE-Transformer: Efficient and Versatile Transformers for Physics Simulations. arXiv preprint arXiv:2505.24717.
How to cite: Whan, K., van der Wiel, K., and Mücke, N.: Emulating transient climate simulations with generative AI , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6079, https://doi.org/10.5194/egusphere-egu26-6079, 2026.
Posters virtual: Thu, 7 May, 14:00–18:00 | vPoster spot 1b
EGU26-19471 | ECS | Posters virtual | VPS23
From bilinear interpolation to machine learning: a comparative assessment of statistical downscaling methods for CMIP6 projections over BrazilThu, 07 May, 14:09–14:12 (CEST) vPoster spot 1b
High-resolution climate projections are essential for climate impact, vulnerability, and adaptation studies, particularly over regions with strong spatial heterogeneity such as Brazil. Although CMIP6 global climate models (GCMs) provide valuable information on future climate change, their coarse spatial resolutions, typically ranging from 100 to 200 km, limit their direct application at regional and local scales. Statistical downscaling techniques offer computationally efficient alternatives to dynamical downscaling, but their relative performance and added value remain insufficiently assessed over Brazil.
In this study, we compare two statistical downscaling approaches applied to a subset of CMIP6 models previously evaluated by Bazanella et al. (2024) – 10.1007/s00382-023-06979-1 – and identified as skillful in representing Brazilian climate: (i) a bilinear interpolation method followed by percentile-to-percentile bias correction, and (ii) machine learning–based downscaling approaches. The original GCM outputs are interpolated to a common high-resolution grid of 10 km × 10 km using bilinear weights, providing a physically consistent reference framework. In parallel, ML-based models are trained using historical GCM predictors and high-resolution reference climate datasets to learn nonlinear relationships and generate high-resolution climate fields.
The performance of both approaches is evaluated for the historical period in terms of mean climatology, spatial patterns, and variability. Future projections under the SSP2-4.5 and SSP5-8.5 scenarios are then analyzed to assess regional climate change signals and associated uncertainties. Results assess the extent to which ML-based downscaling provides added value relative to bilinear interpolation, particularly for variables with strong spatial heterogeneity, such as precipitation and temperature extremes, while also evaluating the ability of the approach to preserve the large-scale climate signals projected by the driving CMIP6 models. This comparative analysis provides insights into the applicability, robustness, and limitations of statistical and ML-based downscaling methods for regional climate assessments over Brazil.
How to cite: Jatobá Santos, D., Goracci, G., Alves Martins, M., and Schneider, R.: From bilinear interpolation to machine learning: a comparative assessment of statistical downscaling methods for CMIP6 projections over Brazil, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19471, https://doi.org/10.5194/egusphere-egu26-19471, 2026.
EGU26-21830 | ECS | Posters virtual | VPS23
Soil Moisture Based Calibration of a Hybrid Hydrological-Neural Network Model in Data Scarce BasinsThu, 07 May, 14:12–14:15 (CEST) vPoster spot 1b
Reliable river discharge simulation generally relies on observed streamflow data for model calibration; however, such observations are often uncertain or unavailable in data-scarce regions, limiting the applicability of conventional hydrological models. This study presents a hybrid modeling framework that uses soil moisture as an alternative calibration variable to improve discharge simulations in the absence of reliable streamflow observations. The framework couples a two-layer version of the daily lumped MISDc (Modello Idrologico Semi-Distribuito in continuo) hydrological model with a Feedforward Neural Network (FFNN), which is employed to enhance parameter calibration by exploiting soil moisture dynamics. The proposed approach is evaluated across three contrasting basins: Tahanaout in semi-arid Morocco, and Colorso (Italy) and Bibeschbach (Luxembourg) in temperate climates. Both in situ and ERA5-Land soil moisture datasets are used as calibration inputs. Model performance is assessed using multiple hydrological metrics, including Mean Absolute Error (MAE), Kling-Gupta Efficiency (KGE), and the correlation coefficient (R). Results show that the hybrid MISDc-FFNN framework substantially improves river discharge simulations compared to the traditional model. Across all basins, MAE is reduced by up to 61%, KGE increases by more than 200%, and R improves by up to 87%, with consistent performance gains observed for both observed and reanalysis-based soil moisture. These findings demonstrate the potential of soil moisture driven calibration strategies to enhance hydrological modeling in data-scarce environments, offering a viable pathway for improved water resources assessment and flood risk management where discharge observations are limited or unreliable.
Keywords: Soil moisture; river discharge simulation; hydrological modeling; machine learning; ERA5-Land; data-scarce regions; feedforward neural network
How to cite: Ait Naceur, K., El Khalki, E. M., Brocca, L., Hadri, A., Jaffar, O., Rachdane, M., Simonneaux, V., Saidi, M. E. M., and Chehbouni, A.: Soil Moisture Based Calibration of a Hybrid Hydrological-Neural Network Model in Data Scarce Basins, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21830, https://doi.org/10.5194/egusphere-egu26-21830, 2026.
