Orals: Thu, 1 May | Room 0.49/50
The early-warning of heatwaves using seasonal forecasting systems has the potential to mitigate economic losses and loss of life. Because of the limited reliability and computational expense of dynamical forecast systems, efforts in recent years have turned to exploiting the power of Machine Learning. Recent years have seen data-driven methods of forecasting deliver added-value for short-term forecasting, yet work on the seasonal scale is not yet as mature. Within the framework of the European Horizon project “CLINT - Climate Intelligence”, a purely data-driven approach to forecasting summer heatwaves on seasonal timescales has been developed. This approach is based on a novel optimisation-based feature selection framework that detects the optimal combination of variables, domains and lag times used to predict heatwaves. The feature selection is performed on multi-millennial paleo-simulation, ensuring sufficient training data, and it is demonstrated that predictors in the model-world are relevant to predictions of the recent past (1993-2016). For forecasts of summer heatwave propensity initialised in May, the data-driven approach matches the skill of the state-of-the-art dynamical multi-model product over Europe, and even outperforms individual systems, at a considerably lower cost. Moreover, low skill over Scandinavia and northern Europe, a long-term issue common to most dynamical systems, is improved in the data-driven approach. Besides forecasts, the data-driven approach also provides insight into the key predictors of European summer heatwave tendency; in particular most-commonly selected predictors correspond to 1-2 months prior to the start of summer (i.e., March) and some have not yet been discussed in existing literature.
How to cite: McAdam, R., Pérez-Aracil, J., Squintu, A., Peláez-Rodríguez, C., Hansen, F., Torralba, V., Loukos, H., Zorita, E., Giuliani, M., Cavicchia, L., Salcedo-Sanz, S., and Scoccimarro, E.: Feature selection for data-driven seasonal forecasts of European heatwaves, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-11905, https://doi.org/10.5194/egusphere-egu25-11905, 2025.
Heatwaves are among the most impactful extreme weather events, posing significant risks to human health, ecosystems, and energy systems. Understanding the return times of these events and assessing how climate change alters their frequency and intensity are critical for effective adaptation strategies. However, the rarity of record-breaking heatwaves in observational datasets makes this task highly challenging. Climate models, while capable of simulating such rare events, require prohibitively long simulations to generate robust statistics for events with return times on the order of centuries.
Our study addresses these challenges by leveraging a dual approach combining rare event simulation algorithms and AI-driven climate model emulators. Rare event algorithms, such as genetic algorithms, efficiently target the extreme trajectories leading to heatwaves while avoiding typical weather conditions, allowing for a more focused exploration of the event space. Although effective for long-duration events, these approaches are less suited to capturing shorter-term phenomena, necessitating novel methodologies for finer temporal scales.
In parallel, we leverage the advancements of deep learning in climate science by training neural networks-based climate model emulators based on Vision Transformers. These emulators drastically reduce computational costs and generate realistic climate simulations, including heatwave dynamics. Here, we explore coupling emulators with a new rare event algorithm specifically designed to sample short and extreme heatwaves. We demonstrate the efficiency of this method by calculating return times for unprecedented heatwave events.
In this work, we use data from PlaSim, a cheap-to-run climate model of intermediate complexity, which enables the verification of return periods spanning up to thousands of years. The next steps involve utilizing more state-of-the-art climate models at finer spatial resolutions and evaluating how the statistics of heatwaves may evolve under various climate change scenarios.
How to cite: Lancelin, A., Wikner, A., Hassanzadeh, P., Abbot, D., Bouchet, F., Dubus, L., and Weare, J.: Coupling AI Emulators and Rare Event Algorithms to Sample Extreme Heatwaves, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-9087, https://doi.org/10.5194/egusphere-egu25-9087, 2025.
High summer temperatures place significant stress on human and natural systems, often leading to severe impacts. Summer hot spells vary widely in terms of intensity and duration, yet event duration is often overlooked or considered a secondary aspect when it comes to studying and predicting such extremes. Different sectors in society, the economy, and the environment are vulnerable to extreme heat on different timescales; therefore, knowing the likelihood of a heat event lasting only a few days or surviving over many weeks is crucial for developing more effective adaptation strategies.
In the last decade, machine learning (ML) techniques have increasingly been used to tackle extreme weather forecasting. Among these, Random Forests (RF) have emerged as an effective tool proven to have some skill in predicting the occurrence and mean amplitude of extreme near-surface temperature events. To the best of our knowledge, such statistical models have yet to be used for the purpose of predicting hot spell duration. This study aims to fill that gap.
The objective of this research is to assess whether a random forest (RF) model can predict the duration of a hot spell from its first day. Specifically, we aim to determine if the model can distinguish between short and long durations, covering both synoptic and subseasonal timescales. To achieve this, we develop a statistical model using data from the Community Earth System Model version 2 Large Ensemble (CESM2-LE) historical runs. For two regions in Western Europe, hot spells are defined as periods when the region-averaged deseasonalised and detrended anomalies exceed 1.5 standard deviations. The model is trained with a number of local and remote predictors, incorporating variables from the land, sea, and atmosphere. These features are provided for the days, weeks and months leading up to the event, as well as for the first day of the event itself.
We perform both a RF classification to predict different duration cohorts (short, medium, long) and a Quantile Random Forest (QRF) to model the full conditional distribution of the response variable (event duration). A key challenge is handling a highly imbalanced dataset, with 3-day events far outnumbering events lasting beyond 10 days.
In addition to shedding light on the statistical and dynamical relationships that drive the persistence of hot spells, the results could be relevant for climate adaptation and policy planning.
How to cite: Pappert, D., Vrac, M., Coumou, D., Tuel, A., and Martius, O.: Predicting Hot Spell Duration with Random Forests, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-17884, https://doi.org/10.5194/egusphere-egu25-17884, 2025.
The rise of deep learning weather prediction (DLWP) models promises to improve short- to mid-ranged weather forecasts out to 14 days. Deep learning models, however, are known in general to perform poorly in conditions that are represented sparsely in the training data and to generalize poorly out of the distribution of the training data. Translated to weather forecasting, this suggests that DLWP models are inaccurate when predicting extreme events that occur only rarely. These extreme events, however, are of highest interest when preventing danger and damage to societies. Here, we therefore inspect how state-of-the-art DLWP models compare to the numerical weather prediction (NWP) model from the European Center for Medium-Ranged Weather Forecasts (ECMWF) on extreme cold and hot spells over North America and Europe. Our results speak not only for DLWP forecasts under normal conditions, but also promise significant skill improvements when forecasting extreme events with DLWP models, emphasized most stongly on cold spells over North America. Similar but weaker trends are observed in cold spell conditions over Europe, as well as in hot spells over North America and Europe. In general, our findings encourage further research in data driven models, such as Pangu-Weather, GraphCast, Aurora, and ECMWF's AIFS. Notably, the advances in DLWP is directly related to decades of research on NWP models. In future research, we will explore the response of DLWP models to warmer climate scenarios that are expected in the later 21st century.
How to cite: Schaible, A., Karlbauer, M., and Butz, M. V.: Deep Learning Weather Prediction Models Exhibit Outstanding Accuracy when Predicting Cold and Hot Spells over North America and Europe, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-11969, https://doi.org/10.5194/egusphere-egu25-11969, 2025.
Accurate air temperature (Ta) forecasting in urban areas is crucial for various socio-economic aspects, including risk warning and optimization of electricity systems. However, forecasting within urban environments faces substantial challenges due to the coarse spatial resolution and inadequate urban representation in numerical weather prediction (NWP) models. In this study, we present a novel multimodal deep learning framework that learns local dynamics from ground-level weather stations while effectively informing large-scale weather patterns for short-range (1- 24 hour lead time) Ta forecasting. The framework first employs graph neural networks (GNNs) to model intra-city spatiotemporal dynamics across 35 weather stations, achieving over 12% forecast improvement compared to modeling individual time series, primarily through mean state regularization. We further develop an end-to-end multimodal framework by integrating the GNN with synoptic weather patterns, achieving an additional 23% improvement, with particular expertise in winter and capturing cold spell events. Our study demonstrates the effectiveness of incorporating multi-scale information from diverse data sources and reveals that weather patterns within approximately 2000 km are critical for local city-scale forecasting. This framework can be readily adapted to other urban areas and will benefit significantly from the increasing deployment of smart IoT sensors to effectively address intra-city temperature heterogeneity.
How to cite: Wang, H. and Yang, J.: Multimodal Deep Learning Framework for Urban Air Temperature Forecasting: Bridging Local and Synoptic Scales, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4940, https://doi.org/10.5194/egusphere-egu25-4940, 2025.
Airliners, struck by lightnings on average once a year, sometimes sustain structural or electrical damage. Even if these incidents generally do not compromise safety onboard due to existing certifications, they lead to costly downtimes and mandatory maintenance operations for the aviation industry. Anticipating the presence of thunderstorm risk areas could help minimize these impacts. Nowadays, predict the exact location of electrical activity in the atmosphere is a complex task because lightning is a non-linear phenomenon which is related to chaotic stormy environments. Numerous variables influence the initiation of electrical discharges, making their modeling using physical equation very challenging. This motivates the use of neural networks to establish relationships between various atmospheric parameters and electrical activity. In the context of aviation safety, this study focuses on the development of a very short term (less than one hour and every five minutes) thunderstorm risk forecasting method above oceans. The proposed methodology is based on computer vision techniques such as neural networks to generate lightning occurrence’s probability maps in the following hour. An encoder-decoder network named ED-DRAP (Che, H et al. 2022) is employed and adapted to the data. In addition to integrating convolutional operations, it also uses spatial and temporal attention mechanisms to process spatio-temporal sequences. Input data come from NOAA’s geostationary GOES-R satellite, including brightness temperature measured by the Advanced Baseline Imager sensor and past electrical activity detected by the Geostationary Lightning Mapper sensor. Outputs from the Numerical Weather Prediction model, Global Forecasting System, are also employed to complement the information provided by satellite imagery. Finally, the model’s outputs are calibrated to produce lightning risk probability maps which are representative of the physical reality, enabling better risk interpretation.
How to cite: Bosc, M., Chan Hon Tong, A., Bouchard, A., and Béréziat, D.: Using deep neural networks for thunderstorm risk prediction., EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-8890, https://doi.org/10.5194/egusphere-egu25-8890, 2025.
We present Prithvi-Typhoon, an innovative adaptation of the Prithvi WxC weather foundation model for tropical cyclone intensity prediction. Through a novel three-stage progressive fine-tuning framework, we bridge the gap between general weather forecasting and specialized tropical cyclone prediction. The model integrates multi-source data from tropical cyclones (1987-2023), incorporating satellite observations, reanalysis products, and historical records. Our architecture features domain-specific feature extraction and multi-scale integration, enabling adaptive balance between local storm features and global atmospheric patterns.
Evaluation results demonstrate substantial improvements over existing methods. Notably, Prithvi-Typhoon shows enhanced skill in predicting rapid intensification events, outperforming both traditional numerical models and existing deep learning approaches. This work represents a advancement in applying foundation models to extreme weather prediction, offering a computationally efficient solution while maintaining physical consistency.
How to cite: Meng, F.: Prithvi-Typhoon: A Foundation Model Approach for Enhanced Tropical Cyclone Intensity Prediction, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-1738, https://doi.org/10.5194/egusphere-egu25-1738, 2025.
Intense cyclones form frequently in the Mediterranean region, with the potential to cause damage to life and property when they hit highly populated coastal areas. Cyclone impacts are caused by the associated strong winds, flash flooding and storm surge. The social and economic impacts are not limited to the Mediterranean area, as cyclones forming in the region can affect Central Europe. While the skill of weather models to forecast such events has dramatically improved over the last decade, the seasonal predictability of Mediterranean cyclones lags behind due to the limitations on horizontal resolution in probabilistic forecasts requiring a large ensemble of simulations. Improving the prediction at a seasonal scale of those extreme events would be of great benefit for society, enabling better disaster risk management and reducing the economic losses they cause. A better prediction of climate extremes would also directly benefit a number of economic sectors such as the insurance and re-insurance industry.
The goal of this work, within the CLINT Horizon project, is to use Artificial Intelligence techniques to enhance the skill of a state-of-the-art seasonal prediction system for predicting Mediterranean cyclones. Here we present results making use of a hybrid AI approach linking the occurrence of those extreme events to their large-scale drivers. The training and validation of different machine learning models is performed using ERA5 reanalysis data. The trained models are then applied to the output of the CMCC operational seasonal forecasts in hindcast mode, and the skill of the modelling chain is assessed. The performance of machine learning models of varying complexity (e.g. random forest, gradient boosting, convolutional neural networks) is evaluated.
How to cite: Cavicchia, L., Ascenso, G., Proserpio, L., Scoccimarro, E., Gualdi, S., Giuliani, M., and Castelletti, A.: AI- enhanced seasonal predictions of Mediterranean cyclones, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15547, https://doi.org/10.5194/egusphere-egu25-15547, 2025.
We present a machine learning based method for predicting extreme precipitation events. This method uses dynamical and thermodynamical variables at coarse resolution as input and the probability of extreme precipitation at higher resolution as the ground truth. Preliminary results show that our detection method, trained on historical EC-Earth3 global climate data and an extreme precipitation mask calculated from the 99th percentile of precipitation from the HCLIM regional model, achieves an accuracy of over 90% for the 2050–2100 period under the SSP126 and SSP370 scenarios within the European domain.
We are working on further improving the method, testing its performance on reanalysis datasets (e.g., ERA5 and CERRA), and adapting it for statistical downscaling and regional climate model emulation.
How to cite: Ivanov, M. and Fuentes Franco, R.: DETEX – Detection of Extreme Precipitation Events in Present and Future Climates at High Resolution Using Machine Learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-12676, https://doi.org/10.5194/egusphere-egu25-12676, 2025.
In recent decades, the prediction of precipitation has become a central focus for atmospheric scientists and weather forecasters. In particular, improving the predictability of rapidly forming rainfall events is critical for protecting lives and property. The island of Ischia, located in the Campania region of Italy, has experienced several landslides and flash floods in recent years with catastrophic effects. To mitigate these geohydrological hazards on this island, we propose a method for short-term rainfall forecasting, with "short-term" defined as a time frame up to six hours. Accurate predictions are essential, as they enable timely implementation of protective measures to safeguard the population.
Accurately predicting rainfall is a complex task influenced by numerous factors, including humidity, temperature, pressure, and wind speed. Historically, rainfall nowcasting has primarily relied on numerical weather prediction (NWP) models. However, this approach has notable limitations, such as high computational requirements and significant processing time, which make NWP models less practical for short-term forecasts.
In the past decade, machine learning (ML) models have revolutionized the way complex problems are addressed and solved, offering solutions that are both fast and highly efficient. Within this domain, deep neural networks (DNNs) a subset of ML have become increasingly prevalent for tackling complex problems using large datasets. Among these, U-Net, a specific DNNs architecture, has proven to be one of the most effective and accurate models for prediction tasks when the input data is image-based. However, achieving high accuracy with such models requires careful preprocessing of the dataset to enhance the model’s ability to effectively learn from the data. Additionally, properly tuning the model's hyperparameters is crucial for optimizing its performance.
In this study, we propose an enhanced U-Net model for nowcasting rainfall with a 120-minute lead time. The input data consists of rainfall radar data and rain gauge measurements. Furthermore, the study evaluates the model's performance under different training scenarios, comparing its efficacy when using only rainfall radar data versus an integrated dataset combining radar and rain gauge data. It is worth noting that the model operates in a regression framework, where the labels or outputs are the rain gauge readings with a 120-minute lead time.
How to cite: Taromideh, F., Santonastaso, G. F., and Greco, R.: Application of a novel deep learning model for precipitation nowcasting , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4735, https://doi.org/10.5194/egusphere-egu25-4735, 2025.
An accurate knowledge of precipitation data at high spatio-temporal resolution is crucial for hydrological forecasting, meteorological analysis, and climate studies. This is especially true in mountainous areas, where traditional climate models struggle to accurately predict precipitation due to factors such as low spatial resolution and where rain gauges are sparse. High-elevation areas are particularly relevant as they act as reservoirs of water resources and are characterized by elevation-dependent climate change signals (Pepin et al., 2022). By leveraging the good performances of the satellite-based IMERG (Integrated Multi-satellitE Retrievals for GPM) rainfall product and the realism of the ERA5 atmospheric reanalysis, we aim to produce a multi-decadal daily rainfall product at the IMERG spatial resolution (roughly 8 km) over the Greater Alpine Region (GAR). To achieve this, we employ advanced machine learning techniques designed to capture the complex, non-linear relationships inherent in atmospheric processes.
Twenty years of IMERG data (from 2001 to 2020) are used to train and test various types of machine learning algorithms to estimate daily precipitation maps starting from some ERA5 atmospheric fields including mid-tropospheric temperature and winds; vertically integrated ice, liquid water and water vapour contents; total precipitation, and other relevant variables. In addition to these atmospheric fields, a high-resolution elevation dataset (ETOPO) is used to represent the intricate terrain of the Alps. The Recursive Feature Elimination (RFE) technique is employed to select key input variables, introducing effective predictors and enhancing the understanding of the influence of physical atmospheric variables and their inter-relationships in mountainous regions. ERA5 total precipitation, vertically integrated ice and water vapour content appear to be the three most relevant input fields for an optimal estimate of IMERG precipitation. Among the algorithms tested (XGBoost, Random Forest, Convolutional Neural Networks, Deep Neural Networks), XGBoost (XGB) is found to be the most reliable and computationally efficient.
The results show a spatiotemporal RMSE improvement of approximately 15 percent, decreasing from 5.18 mm/day (between ERA5 and IMERG) to 4.37 mm/day (between XGB and IMERG). On a seasonal basis, the RMSE is higher in summer and fall, where higher mean precipitation intensities are observed. Also, in terms of changes with the terrain height, the RMSE follows quite tightly the mean precipitation elevation dependence. The XGB model is used to backward extend the IMERG dataset so that precipitation biases and trends can be computed over a multi-decadal time range. These findings demonstrate the potential of machine learning to improve the accuracy of ERA5 rainfall data, which can be exploited to advance our understanding of the emerging elevation-dependent climate change signal.
How to cite: Goudarzi, I., Fazzini, D., Pasquero, C., Meroni, A. N., and Borgnino, M.: A machine learning-based backward extension of IMERG daily precipitation over the Greater Alpine Region, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-369, https://doi.org/10.5194/egusphere-egu25-369, 2025.
The uncertainty quantification in sub-seasonal wind speed forecasting is important for risk assessment and decision-making. One way to improve dynamical forecast skills is to regress information from forecasts of large-scale fields to surface fields by a supervised learning model. For such a statistical downscaling approach, Tian et al. (2024) demonstrated that spatially independent stochastic perturbations based on model residuals can improve the representation of ensemble dispersion. However, this method is limited in fully representing complex spatial correlations and maintaining physical consistency across meteorological fields. Recent advances in probabilistic deep learning models offer promising new approaches for uncertainty quantification, particularly in capturing spatial dependencies.
This study investigates how different statistical downscaling methods can better represent dynamic spatial uncertainty in sub-seasonal ensemble forecasts compared to the independent stochastic perturbation approach. We examine three probabilistic deep learning methods with distinct uncertainty quantification mechanisms: the Quantile Regression for direct modeling of distribution quantiles, the Variational Autoencoders (VAE) for latent space sampling, and the Diffusion model for iterative denoising-based distribution modeling. Our two-stage framework first trains these regression models on the ERA5 reanalysis to establish their capacity for spatial uncertainty representation from the 500hPa geopotential height (Z500) to the surface wind speeds (U100), then applies these probabilistic models to the ECMWF Z500 hindcasts to regress U100 ensembles.
Comprehensive verification reveals distinct characteristics of each method. First, in terms of grid point-wise metrics (the MSE and the CRPS), all these probabilistic methods achieve comparable forecasting skills to independent stochastic perturbations, despite their different approaches to uncertainty representation. Second, spatial structure analysis through Empirical Orthogonal Functions (EOF) analysis and zonal energy spectra demonstrates notable differences: while all methods effectively capture large- and medium-scale features, they differ significantly in representing small-scale spatial correlations. The grid-independent nature of independent stochastic perturbations leads to over-representation of small-scale variations, whereas the Diffusion model shows superior performance across all spatial scales. The Quantile Regression and the VAE show relatively limited skill in capturing small-scale spatial features. These findings suggest that probabilistic downscaling methods, particularly the Diffusion model, can better reconstruct spatial characteristics while maintaining comparable forecasting skills.
Our results indicate that probabilistic downscaling methods can provide more realistic representations of spatial uncertainty compared to the independent stochastic approach, particularly in reconstructing spatial correlations and maintaining physical consistency. This study advances our understanding of how deep learning methods can improve uncertainty quantification in sub-seasonal forecasting.
Tian, Ganglin, et al. "Improving sub-seasonal wind-speed forecasts in Europe with a non-linear model." arXiv preprint arXiv:2411.19077 (2024).
How to cite: Tian, G., Charantonis, A. A., Le Coz, C., Tantet, A., and Plougonven, R.: Improving Spatial Uncertainty Representation in Sub-seasonal Wind Speed Forecasts Using Quantile Regression, VAE and Diffusion, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-11549, https://doi.org/10.5194/egusphere-egu25-11549, 2025.
The quality of the river runoff determines the quality of regional climate projections for coastal oceans or other estuaries. This study presents a novel approach to river runoff forecasting using Convolutional Long Short-Term Memory (ConvLSTM) networks. Our method accurately predicts daily runoff for 97 rivers within the Baltic Sea catchment by modeling runoff as a spatiotemporal sequence defined by atmospheric forcing. The ConvLSTM model predicts river runoff with an accuracy of ±5% when compared to the hydrological model. Compared to more complex process-based hydrological models, ConvLSTM offers fast processing times and easy integration into climate models, demonstrating its potential as a powerful tool for climate simulation and water resource management.
How to cite: Börgel, F., Karsten, S., Rummel, K., and Gräwe, U.: From Weather Data to River Runoff: Using Spatiotemporal Convolutional Networks for Discharge Forecasting, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-12774, https://doi.org/10.5194/egusphere-egu25-12774, 2025.
Climate change is increasing the frequency and intensity of Extreme Weather Events (EWEs), which causes widespread disruption globally. As these events intensify, the need for better hazard identification becomes critical. While machine learning (ML) is already enhancing forecasts, and has huge potential for identifying future hazards. To unlock this potential, we need comprehensive training datasets of historic EWEs that integrate and harmonize diverse datasets, account for data collection discrepancies, and address gaps in temporal and spatial records.
This presentation initially discusses the development of an Aotearoa New Zealand EWE database from 1996 to 2021, which currently includes occurrence data derived from subjective classifications from the national weather service, research organizations, and insurance information. Careful analysis of that database and ancillary reanalyses output can successfully characterise rainfall extreme intensities by deriving duration, peak rainfall, and total accumulation.
Building on that work, this presentation will discuss the development and testing of a methodology to integrate extreme weather event (EWE) occurrence, intensity, and storm track data into a unified database. By processing this combined dataset, we aim to harmonise data from the disparate sources and improve data accuracy and reliability, making it robust for future ML analyses. We also use our experience of applying ML classification schemes in climate research to provide proof-of-concept applications demonstrating the value of our harmonisation methodology.
How to cite: McDonald, A. and Vishwanathan, G.: Developing Extreme Weather Event training datasets to accelerate Machine Learning Applications, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14537, https://doi.org/10.5194/egusphere-egu25-14537, 2025.
The risk of extreme weather under climate change is of paramount importance, but remains one of the most difficult problems to study using conventional physics-based global climate models (GCMs). This is due to the high uncertainty in estimates of extreme weather return times owing to the computational cost of evolving these models for long enough to observe very rare events. AI models trained on historical reanalysis to emulate the dynamics of the global atmosphere have demonstrated both high forecast accuracy and greatly reduced computational cost. Some of these AI emulators can generate stable, decades-long trajectories, which, in conjunction with their affordability, have the potential to greatly reduce extreme weather uncertainties. However, it is impossible to validate if AI emulations can accurately estimate the risk of extreme weather events with return times longer than the historical record. In a first-of-its-kind experiment to assess this capability, we simulate 100,000 years of a stationary climate using PlaSim, a coarse resolution GCM. We then train a selection of stable AI emulators using only 100 years of data, and compare the emulated and true return times of extreme heat waves over Western Europe and the Pacific Northwest. We finally assess how the addition of a land moisture component to these AI emulators improves the accuracy of return time estimates.
How to cite: Wikner, A., Arcomano, T., Lancelin, A., Jakhar, K., Patel, D., Bouchet, F., and Hassanzadeh, P.: Beyond the Unseen: Assessing AI Climate Emulators’ Capacity to Simulate Very Rare Events, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-20607, https://doi.org/10.5194/egusphere-egu25-20607, 2025.
With the development of highly skillful, machine learning-based weather prediction models over the last 2-3 years, many new possibilities have emerged. These include applications, such as downscaling, temporal interpolation, or generating climate storylines, but also a wide range of scientific questions can be (re)examined with the models. One of these is the study of predictability limits by leveraging the full differentiability of the models. For instance, Vonich and Hakim (2024) demonstrated that optimizing initial conditions using the pre-trained GraphCast model significantly reduces forecasting error, even when used with another machine learning-based forecasting model. While this suggests that the improvement in the initial conditions is not only due to compensation in model error, it remains currently unclear to which extent the initial conditions are enhanced by physically meaningful features.
In our work, we aim to address this shortcoming. As a first step, we analyze whether optimized initial conditions can be identified for a broad range of cases by assessing the forecast skill of the model for a larger set of examples. We evaluate the improvement of the forecasts for several variables dependent on the number of optimization steps, the forecast lead time, and for different models. Subsequently, we consider case studies over Europe and compare the optimized initial conditions with data from independent, high quality datasets, in particular local reanalyses and conventional observations. In this way, we examine if the optimized states are physically better aligned with reference data than the original ERA5 initial conditions. To better understand which of the features in the optimized initial conditions lead to the improved forecast, we analyze the null space of the given machine learning-based weather prediction models. This allows us to obtain insight into the information that is exploited by the models for a forecast.
Our work will shed light on the intrinsic predictability limits of weather forecasts and also how MLWP can provide forecasts that outperform equation-based weather prediction models.
How to cite: Brunstein, R., Lessig, C., Rackow, T., and Schlör, J.: Analysis of optimal atmospheric predictability using machine learning-based forecasting models, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13004, https://doi.org/10.5194/egusphere-egu25-13004, 2025.
After numerous successful applications of machine-learning-based global weather models, a new interesting direction of application is to seek high-resolution regional ML-based models that could complement high resolution numerical models serving day-to-day purposes. Development of such a model would combine speed and resource efficiency of ML models with high-resolution capabilities available so far only in the numerical models. Most ML-based models created so far are restricted to the resolution of underlying ERA5 data, often further downsampled due to various constraints, leaving substantial room for further research. With the objective of building a high-resolution ML model for Iceland and equipped with 30 years of 2-km reanalysis data covering Iceland and the surrounding ocean, we are exploring possibilities of the applications of existing ML architectures to our domain. The model we are currently building is based on ClimaX architecture from Microsoft, which we are modifying to best serve our objectives. Understanding the unique needs of regional models during training is one of the key factors in generating a successful regional model. While some of the architectures of the available global models can be applied directly to build a local model, many questions arise: do we need to adjust the cost function during training to handle domain boundaries? Which model levels should we prioritize during training — would it be better to focus on lower levels if the resolution is high and the timescale is short? To what extent can we use transfer learning (leveraging pre-trained weights from the global experiment) and how much will it guide the model toward the optimum? In this talk, we will discuss some of the above considerations for successfully running a regional model and present our high-resolution model for Iceland. The successful development of large machine-learning-based weather models has given weather and climate scientists confidence that models and reanalysis data built over decades are capable of capturing enough variability for ML-based inference. This now opens a new world of possibilities for model improvements and scientific advancements.
How to cite: Stanisławska, K. and Rognvaldsson, O.: Building a high-resolution machine learning weather model, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-19014, https://doi.org/10.5194/egusphere-egu25-19014, 2025.
Recent advancements in AI/ML weather prediction models have attracted significant attention for their innovative approaches to forecasting. These models, leveraging deep learning techniques applied to the European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5 reanalysis data, predict future states of meteorological variables iteratively over specific time steps to generate forecasts. Known as Data-Driven Weather Prediction (DWP), this methodology has demonstrated comparable accuracy to Numerical Weather Prediction (NWP) models for certain variables while requiring substantially less computational effort. Despite its advantages, DWP’s reliance on historical data patterns limits its ability to predict extreme or evolving weather phenomena influenced by global warming and climate change. These limitations present challenges for its application in climate simulations and projections.
To address these limitations, this study explored the application of the GraphCast DWP model in climate research, focusing on global climate downscaling and bias correction. Preliminary experiments with 24-hour GraphCast integrations spanning 36 years (1979–2014) demonstrated that GraphCast’s climate integrations closely align with the mean state and trends of the HiRAM climate simulation. Additionally, the model demonstrates variance in precipitation and surface temperature comparable to ERA5. The primary objective of this study is to demonstrate that this innovative approach to global climate modeling provides both computational efficiency and robust performance, effectively capturing climate phenomena while preserving critical information from climate simulations. Furthermore, the proposed methodology underscores the potential of GraphCast to advance global climate modeling, indicating its suitability for future projections conducted by low-resolution climate models.
How to cite: Tu, C.-Y., Wang, Y.-C., Chou, C.-C., and Yan, Z.-Y.: Application and Evaluation of Data-Driven Weather Prediction (DWP) Model for Climate Modeling, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14049, https://doi.org/10.5194/egusphere-egu25-14049, 2025.
Posters on site: Thu, 1 May, 14:00–15:45 | Hall X5
Flash droughts are a unique natural hazard characterized by their sudden onset and rapid intensification. Accurate and reliable forecasts on subseasonal-to-seasonal (S2S) timescales are crucial for effective preparation and mitigation of the impacts of these events. To enhance the accuracy of soil moisture predictions—a key factor in identifying flash droughts—we propose a hybrid modeling approach that integrates state-of-the-art dynamical forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF) with deep learning techniques (DL).
We use a set of DL models of different complexity for post-processing soil moisture forecasts to not only improve S2S forecasts by correcting systematic errors inherent in numerical weather prediction models, but also to enhance the spatial resolution of the forecasts. This downscaling process is crucial as it addresses a common limitation in S2S forecasts, the coarse spatial resolution that can overlook some variations in soil moisture at a higher spatial scale. By using deterministic inputs, such as the mean and spread from the ensemble forecasting system, we further assess forecast uncertainty through dropout neural networks via Monte Carlo (MC) sampling. This technique allows us to generate probabilistic forecasts by applying MC dropout during the testing phase, thereby generating probabilistic forecasts. Our results show that the DL models outperform the S2S forecasts and lead to skillful S2S forecasts. This advanced modeling framework aims to deliver skillful soil moisture S2S forecasts, ultimately contributing to more effective strategies for managing and mitigating the effects of flash drought events.
How to cite: Otero Felipe, N., Özer, A., and Ma, J.: Deep Learning Postprocessing to Enhance Subseasonal Soil Moisture Forecasts Across Europe, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-168, https://doi.org/10.5194/egusphere-egu25-168, 2025.
We developed a data-driven system for joint prediction of daily precipitation (Pr) and near-surface temperature (T2m) over the global domain by utilizing NASA’s satellite observations and the associated reanalysis products, with the focus on S2S hydrologic forecasting. Our approach is based on a well-established methodology of linear inverse modeling modified and adapted by our science team for high-resolution modeling of precipitation. The key element of this new methodology is the usage of a so-called pseudo-precipitation (PP) variable, equal to the actual Pr where precipitation is occurring and, otherwise, equal to the (negative) air-column integrated water-vapor saturation deficit — the amount of water vapor to be added to the air column to achieve saturation at each vertical level. The model’s jointly obtained Pr and T2m forecasts are then validated against the observed fields as usual.
The above model is shown to be an efficient tool for emulating daily sequences of global coupled T2m and Pr fields with spatiotemporal characteristics strikingly similar to the observed characteristics. We used a large (100-member) ensemble of our statistical model’s hindcasts of precipitation over global domain to predict probabilities of weekly and biweekly precipitation amounts in one of the three categories (below normal, normal, and above normal) and compared these hindcasts with those based on the NASA GEOSS2S v2p1 model (4-member ensemble), calibrated using extended logistic regression. While the statistical model’s S2S precipitation forecast skill is somewhat lower than that of the reference NASA state-of-the-art system, it exhibits similar geographical and seasonal distributions, which warrants further research. We are currently looking into incorporating automated ML/AI feature identification techniques into our existing set up (with a linear activation function), to fine-tune the model learning and improve its predictive potential.
How to cite: Kravtsov, S., Robertson, A., Yuan, J., and Ghadamidehno, M.: Emulation and S2S probabilistic prediction of 2-m temperature and precipitation over the global domain using linear inverse modeling, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-3286, https://doi.org/10.5194/egusphere-egu25-3286, 2025.
Precipitation nowcasting is a long-standing challenge due to the inherent unpredictability, which often lead to significant risks and damage. While traditional approaches focus on modeling the nonlinear relationship between initial precipitation states and future states, these methods often fail to capture accurate precipitation dynamics, such as its distribution and intensity. The absence of guidance from physical theory limits data-driven methods in disclosing the chaotic nature of precipitation. To address this, we integrate Prandtl’s mixing length theory from fluid dynamics with diffusion models commonly used in computer vision to enhance the prediction of precipitation distributions and details over the next 200 minutes. This integration accounts for the turbulent properties of precipitation, improving both accuracy and granularity in forecasts. Additionally, we leverage multi-source data, particularly lightning observations, to train a control network for our diffusion model. This enhancement allows for more accurate and controllable predictions of precipitation initiation, decay, and overall spatial-temporal patterns. Our approach advances the state of the art in precipitation nowcasting, offering a robust framework that bridges physical theory with modern deep learning techniques.
How to cite: Li, D., Deng, K., Zhang, D., Leng, H., Ren, K., and Song, J.: Precipitation nowcasting diffusion model based on turbulence theory and multi-source data, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-5319, https://doi.org/10.5194/egusphere-egu25-5319, 2025.
While short-term weather forecasting has benefited from extensive data and research, leading to high predictive accuracy, long-term forecasts, particularly medium-range predictions, lag significantly due to data scarcity. This research aims to bridge this gap by leveraging the advancements in Artificial Intelligence (AI), particularly Deep Learning. We propose a novel approach using Neural Ordinary Differential Equations (NODEs), which represents a transformative step in dynamic systems modeling. Neural ODEs offer a flexible and powerful framework for continuous-time models, which is particularly beneficial for handling sparse or irregularly sampled data prevalent in climate studies. Our methodology utilizes the Empirical Orthogonal Function (EOF) to extract principal component time series from limited climate data. These components serve as inputs for NODEs to predict future climatic conditions. This approach is innovative in its ability to handle non-linearities and temporal dependencies in climatic data, making it highly suitable for medium-range weather forecasting. The potential of NODEs in this context is significant, as they provide a means to accurately predict weather patterns with less data, a common limitation in long-term forecasting. By enhancing the precision of medium-range forecasts, this research contributes to more effective climate change adaptation and mitigation strategies, ultimately aiding in the safeguarding of ecosystems and human societies against the adverse effects of extreme weather conditions.
How to cite: Lee, J. and Moon, W.: Subseasonal to Seasonal Forecast Using Neural Ordinary Differential Equations, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7861, https://doi.org/10.5194/egusphere-egu25-7861, 2025.
Wind gust is a sudden meteorological weather phenomenon. It can cause many material and moral accidents, especially if it occurs during aircraft take-off and landing at airports. In this study, gust analysis and gust prediction for Istanbul Airport were performed using machine learning algorithms. Metar data of Istanbul Airport between 01.11.2018 and 31.12.2024 were used in the study. When this Metar data was analysed, it was found that on average between 250 and 300 Gust events were reported annually. Gust values were found to vary between 11 and 65 knots. It was reported that the highest number of gust events was reported in November with 179 times and the lowest number was reported in August with 38 times. When the gust intensities are analyzed, it is seen that the strongest gusts occurred in February. When the gusts were analyzed hourly, it was found that most gusts occurred between 01.00 and 03.00 hours. The most severe gusts occurred between 15.00 and 20.00. In the study, the relationship between gusts and other meteorological variables such as temperature, pressure, dew point temperature was analyzed. In the other part of the study, three different machine learning methods Random Forest (RF), long-short term memory (LSTM) and extreme gradient boosting (XGB) were used to predict gusts. In these methods, models were derived and evaluated on 1000 different randomly selected subsets, 70% for training and 30% for testing. It was observed that the prediction success of the three different models used in the study increased at times of high wind gust values (≥ 30 knots), while the prediction success was lower at times of low wind gust values.
How to cite: akbayır, I., yavuz, V., Demirhan, D., and İnanç, B. M.: Meteorological Analysis and Prediction of Gusts at Istanbul Airport Using Machine Learning Algorithms, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-12021, https://doi.org/10.5194/egusphere-egu25-12021, 2025.
As part of the OptimESM project, this work aims to prototype a framework for downscaling post-CMIP6 Earth System Models (ESMs) to refine long-term projections up to 2300. This effort focuses on understanding regional climate impacts and extreme events, including heatwaves, droughts, and precipitation extremes, with the goal of supporting robust regional climate projections and informing adaptation strategies across Europe. Within this broader context, our study investigates the application of deep learning techniques to downscale daily temperature fields, enhancing the detection and characterization of European heatwaves through improved spatial resolution. Utilizing the open-source DeepR library based on Transformer architecture, we obtained a five-fold downscaling from ERA5 to CERRA datasets. Performance evaluation highlighted significant improvements in detecting heatwaves, particularly in mountainous areas. Integrating high-resolution orography data increases accuracy by 53%, improving the detection rates of heatwave days from 18% (ERA5) to 27% (DeepR) in regions like southern Norway during the validation period 2015-2020. Despite some perceptual improvement, challenges remain in generalizing across spatial domains and accurately modeling temperature distribution tails, which are critical for extreme events. To address these limitations, we explore advanced architectures such as UNet and Diffusion Models, alongside high-resolution land-cover data and enhanced land-sea masks.
How to cite: Tian, T., Ronzani, H., Beauchamp, M., Su, J., Krus, K., Yang, S., and Fuentes-Franco, R.: Enhancing European heatwave characterization: deep learning-based downscaling of global climate data, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13581, https://doi.org/10.5194/egusphere-egu25-13581, 2025.
Weather forecasting plays a critical role in preventing natural disasters and improving convenience in daily life. However, traditional physics-based numerical weather prediction models have limitations in real-time and high-resolution predictions due to computational complexity and restricted computational resources. This study aims to enhance the predict skill of short-term weather forecasting by utilizing deep learning technologies. Particularly, this study attempts to seek developing methodologies to improve the skill of short-term rainfall forecasts produced by the Korea Meteorological Administration through artificial intelligence. By addressing systemic biases and errors in rainfall prediction data, this research aims to enhance predictive performance. Weather forecast data collected at 1-hour intervals—including temperature, wind speed, humidity, and precipitation—was preprocessed and used as input for the deep learning model. A deep neural network-based architecture was designed for building the forecast model. The model was trained, validated, and evaluated using data spanning the past three years. This study is expected to improve the skill of short-term weather forecasts while enhancing computational efficiency compared to conventional physics-based numerical weather prediction models. Furthermore, the proposed model demonstrates high potential for application in various fields, including disaster management, agriculture, and energy management.
How to cite: Jang, S., Shin, J.-Y., Park, J., Kim, S., and Lee, G.: Development of a Deep Learning-Based Weather Forecasting Model Using Short-Term Neighborhood Forecast Data, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14190, https://doi.org/10.5194/egusphere-egu25-14190, 2025.
Weather forecasting is crucial for human activities, yet traditional numerical models often face limitations due to complex physical processes and high computational cost. Deep learning–based neural networks offer a promising alternative. The Spherical Fourier Neural Operator (SFNO) model introduces the Spherical Harmonic Transform to maintain SO(3) rotational invariance, ensuring long-term stability in forecasts and preventing early collapse. However, we have identified two key shortcomings in SFNO: high memory consumption and limited ability to capture high-frequency information due to the truncated of spectrum.
To address these issues, we propose the SFF model, which improves upon the well-known SFNO model primarily in the following ways:
- a) U-Structure: We add up-sampling and down-sampling operators between SFNO blocks, allowing the initial and final stages of the SFNO block chain to handle broader frequency spectra, while the middle layers focus on relatively low-frequency information. Under a limited memory budget, this design enables us to increase the number of SFNO blocks or enlarge the embedding dimension, thereby enhancing forecast accuracy.
- b) Vision Transformer-like Residual Connection: We introduce a Vision Transformer–like architecture between the encoder and decoder as the skip connection, and specialize it to focus on local features. This strengthens the model's ability to capture high-frequency information, enhances its capacity for local feature learning, and leads to more robust and accurate predictions.
Considering the discontinuous occurrence and development of precipitation, SFF employs an independent precipitation model which can be easier to learn the physical processes of precipitation and leverages classification weighting to improve the detection and prediction accuracy of heavy rainfall, further extending the effective lead time of precipitation forecasts through joint training.
We conducted experiments on ERA5 dataset, using data from 1979–2017 for training, 2018 for validation, and 2020 for testing. The experiment results demonstrate that SFF can generate stable 30-day forecasts cost-effectively on a single NVIDIA H20 GPU, with key metrics—such as the root mean square error (RMSE) and anomaly correlation coefficient (ACC) for Z500/t2m/t850 comparable to the well-established IFS model, and better than the SFNO model. Meanwhile, for precipitation predictions, SFF also exhibits a forecast skill level comparable to that of the IFS model. Moreover, for heavy rainfall prediction, SFF achieves a Threat Score (TS) of over 0.25 in single-step forecasts for 70 mm of precipitation. After joint training of SFF and the precipitation model, the precipitation score within 10-day forecasts can be improved by 5% compared to direct coupling. This study underscores the potential of Neural Operator–Based AI models in advancing weather forecasting and extreme weather prediction.
How to cite: Jia, Q., Dai, Z., Wang, C., Au Yeung, I., Jing, H., Zhang, R., Sun, J., and Xue, W.: Sphere Fusion Forecast (SFF): A Neural Operator–Based Model for Global Weather Forecasting, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-18929, https://doi.org/10.5194/egusphere-egu25-18929, 2025.
Agricultural drought (AGD), defined by a deficit in soil moisture, is a complex natural hazard phenomenon that causes extreme damage to water supply, food production, and socio-economic loss at different time scales. India is a developing country, and 60% of its population depends on agriculture. India has experienced frequent extreme drought conditions in the last few decades, for example, the 2015-16 North Indian and 2017-18 Southern Indian drought, where more than 330 million people were affected due to food unavailability and shortage in groundwater resources. The spatial patterns of AGD vary significantly in India due to uncertainty in regional climatic conditions caused by the immense increase in global warming. The prediction of agricultural drought at a sub-seasonal scale would help the farming community to plan appropriate crops for the season and conserve water for irrigation.
This study proposes a statistical framework to predict the agricultural drought with 1-, 2-, and 3-month lead times over the Indian subcontinent. Soil moisture percentiles (SMP) are utilised as a drought index where values less than 20th percentiles represent drought conditions. SMP is a widely used drought index in research because it directly represents the water content in the soil and responds relatively quickly to changes in soil water content due to variations in rainfall and irrigation. The variation of SMP depends on various hydroclimatic parameters at local and non-local scales. Thus, this study has considered the air temperature (max. and min.), Potential Evapotranspiration, Vapour Pressure Deficit, Rainfall, soil moisture percentile, Normalised Difference Vegetation Index, El-Nino southern oscillation, North Atlantic Oscillation, Indian Ocean Dipole, Pacific Decadal Oscillation, and Madden Julian Oscillation as a predictor (or feature) from the various satellite (NOAA-19, 20, and AVHRR) and observational (IMD – Indian Meteorological Department) data sources. The Long-Short-Term Memory (LSTM) model, with an MSE custom loss function, is used to forecast agricultural drought. The model was trained from June 1981 to May 2015 and tested at each grid point cell between June 2015 and May 2022. The model performance is examined using Pearson’s correlation > 0.6 for a 1-month lead and further decreased for a 2 and 3-month lead. The forecasting matrices such as percentage porrect, POD, FAR, and ETS indicated that the predictability of AGD is comparably high over northern, southern, and north-eastern India. At last, the trained models are used to discover variables that, depending on feature relevance, influence agricultural drought predictability on a sub-seasonal scale. The result shows that vapour pressure deficit followed by maximum temperature, Pacific decadal oscillation, and soil moisture percentile are the primary features that control drought predictability.
How to cite: Verma, S. and Lanka, K.: Sub-seasonal Prediction of Agricultural Drought in India Using Long-Short-Term Memory Networks, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15750, https://doi.org/10.5194/egusphere-egu25-15750, 2025.
The generation of accurate and reliable short-term forecasts (<12 hours) of near-surface (~10 m above ground level) gridded wind speed data, hereinafter called NSWS, are crucial for various socioeconomic and environmental applications. For instance, in the face of climate change, accurate wind speed predictions can contribute to the decarbonization of the electricity grid by optimizing the wind energy generation.
Traditional NSWS forecasting methods relies on Numerical Weather Prediction (NWP) models, which require significant computational resources, particularly when high spatial and temporal resolution are required. Moreover, these models often yield inaccurate results, especially in regions with complex topography. As a more efficient alternative to this pressing issue, the Climatoc-Lab, as part of the PTI+Clima, is exploring Artificial Intelligence (AI) methods to enhance the efficiency and accuracy of short-term NSWS predictions. We propose the use of two deep learning methods:
-
A U-Net architecture based on Partial Convolutions to generate high-resolution hourly NSWS maps from station-based observations.
- An encoder-decoder architecture based on mixed convolutional and recurrent (ConvLSTM) layers to predict short-term NSWS maps using the generated infilled data as input.
This AI-based product, designed as an early warning system, generate high-resolution (~3/9-km) short-term (12 h; 1-h resolution) NSWS forecasts in near real-time (seconds) using a GPU.
Measurements from meteorological station networks provide accurate site-specific observations, capturing local wind effects, but with limited spatial coverage, being sparse and almost absent in mountainous and remote areas. Conversely, reanalysis and simulation products offer complete spatial coverage at low resolution but fail to accurately reproduce local NSWS. Our AI-based tool combine the strenghts of both worlds, as it is trained using both, observation and simulation data. The observations are provided by the Spanish Meteorological State Agency (AEMET), while the simulation data comes from reanalysis like ERA5-Land (9-km).
The AI-based tool achieves a high correlation of 0,96 for Infilling and 0,849 for Prediction for the year 2020 of ERA5-Land data used for validation, with potential for further improvements. This also shows a reasonably high correlation of 0,84 with the AEMET meteorological observations. This scalable AI-based approach promises to enhance short-term NSWS forecasting for AEMET and other meteorological services, highlighting the promising role of AI to improve both forecast precision and operational efficiency in meteorology applications.
How to cite: Martínez-Roig, M., Plaza-Martín, N. P., Azorín-Molina, C., Monsalvez-Pozo, K., Andrés-Martin, M., Chen, D., Zeng, Z., Vicente-Serrano, S. M., McVicar, T. R., Guijarro, J. A., and Safaei-Pirooz, A. A.: AI-based Short-Term Wind Speed Forecasting for Real-Time Applications., EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-19963, https://doi.org/10.5194/egusphere-egu25-19963, 2025.
Earth system models (ESMs) play a vital role in understanding and forecasting the dynamics of the Earth's climate system. Accurate simulation of precipitation is especially critical for evaluating the impacts of anthropogenic climate change, anticipating extreme weather events, and devising sustainable strategies to manage water resources and mitigate related risks. However, ESMs often exhibit significant biases in precipitation simulation due to the wide range of scales involved in these processes and the substantial uncertainties they encompass. Moreover, due to computational constraints, ESM simulations still have low horizontal resolution compared to the scales relevant for precipitation.
In this work, we present a novel framework to improve the representation of precipitation in ESMs by integrating physically modeled circulation variables with state-of-the-art generative diffusion models. Based on large-scale (1 degree) circulation fields, our method produces accurate high-resolution (0.25 degree) precipitation estimates at global scale. Our approach introduces stochasticity into the precipitation field, significantly improving the representation of extreme events and fine-scale variability while maintaining the fidelity of large-scale patterns. Our proposed methods thus provides an alternative to traditional column-based parameterization, avoiding the need for a posteriori bias correction and downscaling.
Preliminary results highlight the ability of our generative model to produce precipitation fields with substantially smaller biases compared to those derived from classical parameterizations of the GFDL model, while achieving higher spatial resolution. In future climate scenarios, precipitation derived from parameterizations often becomes increasingly uncertain, whereas circulation variables, being more directly tied to large-scale dynamics, may provide a more stable foundation for generating high-resolution precipitation fields. Building on this, we demonstrate the application of our framework to generate daily high-resolution precipitation maps for future climate projections, offering an improved and robust tool to address critical challenges in climate impact studies.
How to cite: Aich, M., Bathiany, S., Hess, P., Huang, Y., and Boers, N.: Uncertainty-aware precipitation generation for Earth system models with diffusion models, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7021, https://doi.org/10.5194/egusphere-egu25-7021, 2025.
The increasing frequency and severity of extreme weather events, such as heavy rainfall and flooding, emphasize the urgent need for advanced early warning systems. Short-duration rainfall extremes, exacerbated by climate change, significantly increase flood risks, particularly in urban coastal cities like Mumbai. Mumbai's vulnerability arises from rapid urbanization, its coastal location, and variable topography, which contribute to significant spatial variability in rainfall. We have used Global Forecast System (GFS) data to identify key predictors for high-resolution, 3-hour rainfall forecasts for Mumbai. The GFS variables were selected using a correlation matrix. We have used past 3-hour observed rainfall data from Automatic Weather Stations (AWS) across 15 locations in Mumbai (2015–2023) along with selected GFS variables, which include Precipitable Water, Precipitation Rate, Relative Humidity, and Total Cloud Cover, to forecast rainfall for one timestep ahead. The dataset was divided into 80% for training and 20% for testing. We employed a hybrid Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) model to enhance forecast accuracy. The CNN captures spatial features, while the LSTM models temporal dependencies, effectively addressing the challenges of hyperlocal rainfall forecasting. Further, we incorporated a weighted Mean Squared Error (MSE) loss function to prioritize extreme rainfall events (≥95th percentile). The results indicate that using CNN-LSTM models reduced the Root Mean Square Error (RMSE) by 9.41% -12.38% and increased the Correlation Coefficient (CC) by 70.4%-113% compared to GFS models. At the 95th percentile, the Hit Rate (HR) improved by 233% -483.3%, while the False Alarm Rate (FAR) decreased by 7%-16.2%. Using weighted MSE also enhanced performance, increasing the HR by 255.5%-583.3% at the 95th percentile and reducing the FAR by 7% -13.2%. Implementing weighted MSE as a loss function resulted in a reduction in RMSE by 9.94% -12.86% and an increase in CC by 85.2%-126%. This study highlights that the hybrid CNN-LSTM model, combined with a weighted MSE loss function, demonstrates superior capability in accurately forecasting 3-hourly extreme rainfall events in Mumbai, providing critical advancements for early warning systems and flood risk mitigation.
How to cite: Tripathy, P., Murtugudde, R., Karmakar, S., and Ghosh, S.: Enhancing Hyperlocal 3-Hourly Rainfall Forecasting for Mumbai Using a Hybrid CNN-LSTM Model., EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-1037, https://doi.org/10.5194/egusphere-egu25-1037, 2025.
AI weather models are becoming valuable tools for predicting the weather. While AI models’ general forecasts are known to be skillful, their forecast skill of extreme events is not fully understood. The 2021 Pacific Northwest (PNW) heatwave is a good case study for AI models because it falls outside of the distribution of heat waves in AI model training datasets.
Here, we investigate the forecast performance of 8 AI models (AIFS, Gencast, NeuralGCM, Graphcast, Fuxi, Pangu, Fourcastnet, FourcastnetV2) of the 2021 PNW heatwave. Despite the event being out of the training dataset distribution, their forecast performance is comparable to that of a state-of-the-art numerical weather prediction model (IFS). Specifically, AI models and IFS can accurately forecast the heatwave for lead times less than 7 days.
Two recent studies suggest the predictability barrier of the PNW heatwave may be due to an initial condition observation error. Leach et al. (2024) found that the 26th ensemble member of a 250 member IFS forecast accurately forecasts the heatwave 12 days in advance. Vonich and Hakim (2024) used backpropagation in Graphcast to find an optimal initial condition that leads to an accurate forecast 10 days in advance. Are these initial conditions robust across an ensemble of AI models? And do these initial conditions point to a unique solution?
We find a large spread in forecast accuracy when running the 8 AI models with the Leach et al. (2024) and Vonich & Hakim (2024) initial conditions. Furthermore we ran 1000 member ensembles in NeuralGCM and find initial conditions that lead to an accurate long-term forecast are not unique. These results suggest that the improvement in forecast accuracty to certain initial conditions may not necessarily be due to the initial conditions being closer to ground truth but rather they are due to cancelation of model error.
How to cite: Miyawaki, O., Fei, C., Li, S., Patel, D., Sarro, G., Zhang, H., Marchakitus, A., Hassanzadeh, P., Abbot, D., Weare, J., Nakamura, N., and Shaw, T.: On the robustness of AI model forecast skill and initial condition uncertainty of the 2021 Pacific Northwest Heatwave, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4576, https://doi.org/10.5194/egusphere-egu25-4576, 2025.
AXA is developing in-house Natural Hazard models (or Natural Catastrophe models) in order to gain a deeper understanding of, evaluate, and monitor the climate risks underpinning (re)insurance portfolios. Such models simulate large numbers of synthetic weather events to estimate the probability of rare and extreme events, enabling more robust risk management and informed decision-making.
AI-driven weather models offer the capability to rapidly produce thousands of unique ensemble scenarios of low-likelihood high-impact weather events such as tropical cyclones. This study specifically utilizes tropical cyclones (TCs) as a primary illustration of the potential of AI-based weather models for risk management.
In this study we use FourCastNet SFNO, the global data-driven weather forecasting model developed by NVIDIA available on the NVIDIA Earth-2 platform to simulate historical but also synthetic (i.e. never observed) hurricanes. SFNO trained on ECMWF ERA5 reanalysis data provides short to medium-range global predictions at 0.25° resolution. A large ensemble of hurricane simulations is performed using the HENS method, developed at Berkeley, the NVIDIA leveraging Earth2Studio from NVIDIA’s Earth-2 platform.
HENS-SFNO performance is first assessed by evaluating the model's ability to reproduce post-2017 historical hurricanes (intensity, track, landfall location). HENS-SFNO capabilities in simulating synthetic hurricanes are then assessed in a second step by evaluating track density and landfall frequencies by categories of hurricanes against the historical tropical cyclone IBTrACS database.
How to cite: Meynadier, R., Renard, X., Koch, M., Rakotoarimanga, H., Ertl, G., Leinonen, J., and Detyniecki, M.: Use of NVIDIA FourCastNet model to improve tropical cyclones risk modelling. , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7075, https://doi.org/10.5194/egusphere-egu25-7075, 2025.
Accurate high-resolution spatio-temporal weather forecasting is vital for advancing our understanding of regional weather dynamics and improving meteorological applications. Traditional forecasting relies on numerical weather prediction (NWP) models, which are computationally demanding, particularly when implemented for large domains and high-resolution grids. Recently, Deep Learning (DL) has emerged as a powerful alternative, leveraging historical data to identify patterns and predict future atmospheric conditions. In this work, we develop a regional DL-based forecasting system tailored for the Arabian Peninsula (AP), a region with unique climatic conditions characterized by extreme temperatures and high wind energy potential. Therefore, it serves as an ideal case study for regional weather forecasting. The developed system forecasts hourly meteorological variables such as wind speed, wind direction, and temperature at a 5 km spatial resolution up to 48 hours ahead, with a focus on key vertical levels relevant to wind energy applications. Two forecasting approaches are explored: recursive forecasting, which iteratively advances fine-scale spatio-temporal states over time, and downscaling, which refines coarse-resolution forecasts of the meteorological variables into their high-resolution counterparts. Additionally, we propose a combined approach that integrates these methods by combining fine-scale dynamics propagation with coarse-scale to fine-scale refinement. The frameworks were evaluated both qualitatively and quantitatively, demonstrating that while recursive forecasting accumulates errors over time, the downscaling approach effectively produces high-resolution forecasts. The combined approach significantly improves the forecasting precision, offering robust performance at early time steps and reduced error accumulation over extended forecasting horizons.
How to cite: Resifi, S., Al Aawar, E., Dasari, H., Jebari, H., and Hoteit, I.: Regional High-Resolution Weather Forecasting over the Arabian Peninsula: A Data-Driven Approach, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-1879, https://doi.org/10.5194/egusphere-egu25-1879, 2025.
Predicting climate extremes such as droughts, heatwaves, and heat stress episodes remains a critical challenge in Earth system sciences. Current state-of-the-art methods often fail to deliver reliable forecasts, especially at subseasonal-to-seasonal (S2S) timescales (i.e., from two weeks to two months in advance). As global climate variability continues evolving, the need for advanced, trustworthy, data-driven forecasting methodologies has never been more pressing.
Extended numerical weather prediction systems, such as those led by the European Centre for Medium-Range Weather Forecasts (ECMWF), remain the primary method for S2S prediction (Vitart & Robertson, 2018). While recent deep learning approaches have demonstrated remarkable competitive performance (e.g. Olivetti & Messori, 2024), proposed models predominantly focus on global-scale average weather predictions, overlooking critical local-scale extreme events (Pasche et al., 2024). Moreover, creating accurate probabilistic forecasts conditioned on the initial state remains a significant challenge within the scientific community. In the context of weather forecasting, traditional statistical methods, such as ensemble-based techniques that generate multiple forecasts to estimate uncertainty, are commonly used. These approaches include techniques such as introducing noise into initial states, varying neural network parameters, or training generative models. While generative models offer the most robust solutions, they demand substantial computational resources and extensive data availability.
Here, we evaluate several state-of-the-art dynamical weather forecasting systems, such as those of ECMWF and the National Centers for Environmental Prediction (NCEP), together with recently-proposed deep learning models on their ability to predict extreme heatwaves across all continents at S2S timescales. Since uncertainty quantification is essential for supporting practical decision-making, we focus on deep learning models that provide probabilistic forecasts and have publicly available source code. These include FourCastNet, proposed by Kurth et al. (2023), as well as ArchesWeather and ArchesWeatherGen, developed by Couairon et al. (2024). This analysis underscores the limitations of contemporary deep learning and dynamical weather forecasting systems in reliably and probabilistically predicting S2S extremes, while also providing a valuable benchmark to guide future research efforts.
References:
Couairon, G., Singh, R., Charantonis, A., Lessig, C., & Monteleoni, C. (2024). ArchesWeather & ArchesWeatherGen: a deterministic and generative model for efficient ML weather forecasting. arXiv preprint arXiv:2412.12971.
Kurth, T., Subramanian, S., Harrington, P., Pathak, J., Mardani, M., Hall, D., Miele, A., Kashinath, K., & Anandkumar, A. (2023). FourCastNet: Accelerating global high-resolution weather forecasting using adaptive Fourier neural operators. Proceedings of the Platform for Advanced Scientific Computing Conference (PASC '23), Article 13, 1–11. https://doi.org/10.1145/3592979.3593412
Olivetti, L., & Messori, G. (2024). Do data-driven models beat numerical models in forecasting weather extremes? A comparison of IFS HRES, Pangu-Weather, and GraphCast. Geoscientific Model Development, 17(21), 7915-7962.
Pasche, O. C., Wider, J., Zhang, Z., Zscheischler, J., & Engelke, S. (2025). Validating Deep Learning Weather Forecast Models on Recent High-Impact Extreme Events. Artificial Intelligence for the Earth Systems, 4(1), e240033. https://doi.org/10.1175/AIES-D-24-0033.1
Vitart, F., & Robertson, A. W. (2018). The sub-seasonal to seasonal prediction project (S2S) and the prediction of extreme events. npj climate and atmospheric science, 1(1), 3.
How to cite: Decancq, C., Mortier, T., Hagan, D., Deman, V., Insua Costa, D., Camps-Valls, G., Coumou, D., and Miralles, D.: Benchmarking Deep Learning Models for Probabilistic Subseasonal Forecasting of Heat Extremes, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-9541, https://doi.org/10.5194/egusphere-egu25-9541, 2025.
Posters virtual: Mon, 28 Apr, 14:00–15:45 | vPoster spot 5
EGU25-10613 | ECS | Posters virtual | VPS5
FarmD: A Web Interface for Visualization of Predicted Weather Parameters Using 1D Transformer Hybrid ModelsMon, 28 Apr, 14:00–15:45 (CEST) | vP5.10
Precisely predicting weather parameters is crucial for precision horticulture, especially in horticultural lands where timely environmental insights significantly impact crop yield and quality. This study presents a novel hybrid modeling approach employing 1D Transformer networks integrated with traditional machine learning techniques to predict hourly temperature variations. Utilizing the ERA5 reanalysis dataset spanning from 1940 to December 2024, the hybrid model efficiently captures location-specific spatiotemporal dependencies and nonlinear trends in historical weather data.
The predicted weather data generated by the hybrid model is used in FarmD, a web-based user interface developed for farmer-centric applications. FarmD provides real-time visualization of critical weather parameters, including temperature, relative humidity, wind patterns, rainfall, and soil temperature, specifically tailored to horticultural regions. Through its intuitive interface, users can query predicted and historical data by selecting attributes, dates, and times, with an option for location-specific searches to support targeted agricultural decision-making.
This integration of predicted data with an accessible web platform highlights significant advancements in delivering actionable insights to end users. By combining advanced computational methods with user-focused design, FarmD enables horticulturists to adopt data-driven practices, contributing to sustainable and efficient agricultural management.
How to cite: Ramalingam, S.: FarmD: A Web Interface for Visualization of Predicted Weather Parameters Using 1D Transformer Hybrid Models, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-10613, https://doi.org/10.5194/egusphere-egu25-10613, 2025.
EGU25-14457 | Posters virtual | VPS5
Evaluating the extrapolation capability of deep learning in rainfall-runoffMon, 28 Apr, 14:00–15:45 (CEST) | vP5.11
Deep learning, a prominent artificial intelligence method, is increasingly applied in research addressing the impacts of global warming in the future. However, it is widely acknowledged that deep learning exhibits limitations in extrapolation, as it typically predicts accurately only within the range of the training data. When future scenarios extend beyond this range, the reliability of predictions can diminish significantly. In Japan, for example, the annual maximum precipitation is reported to be increasing, according to the Japan Meteorological Agency, indicating a potential for future values to exceed historical records. Despite this, limited studies have explored the extent to which deep learning methods can reliably extrapolate beyond the training data range. This study quantitatively evaluates the extrapolation capability of deep learning in hydrology, specifically focusing on rainfall-runoff modeling at the watershed scale. Meteorological data, including precipitation and temperature, are utilized as inputs, while river flow serves as the output. The Long Short-Term Memory (LSTM) model, which is well-suited for time-series data, was employed as the deep learning framework. Data were partitioned into training, validation, and test datasets, with river flow values categorized using threshold percentiles of 90, 95, 97, 98, and 99, rather than conventional time-based splits. This approach allows for a focused investigation into the range of accurate extrapolation beyond the training dataset. Preliminary findings reveal that the LSTM model successfully captured peak river flows up to 250.1% higher than the maximum values of the observed river flow discharge in the training-validation dataset. These results demonstrate the potential for deep learning to extrapolate in hydrological modeling, though further research is necessary to assess the performance of alternative deep learning methods and additional case studies.
How to cite: Junsei, S.: Evaluating the extrapolation capability of deep learning in rainfall-runoff, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14457, https://doi.org/10.5194/egusphere-egu25-14457, 2025.
EGU25-1959 | Posters virtual | VPS5
A Comparative Analysis of Data-Driven Machine Learning Models for Rainfall Forecasting in BangladeshMon, 28 Apr, 14:00–15:45 (CEST) | vP5.12
Accurate rainfall forecasting is crucial for effective urban planning and disaster management in Dhaka, the capital of Bangladesh, a city highly vulnerable to urban flooding and extreme weather events. Traditional forecasting methods often struggle to capture the region's complex rainfall patterns, resulting in inaccurate rainfall forecasts. This study evaluates the performance of two traditional machine learning algorithms, Random Forest Regression and Multi-layer Perceptron (MLP), alongside one deep learning algorithm, the Long Short-Term Memory (LSTM) network. These models are trained and tested to forecast rainfall over 1 to 5-day lead times, emphasizing their ability to handle temporal dependencies in time series data. Atmospheric and hydrologic variables, including temperature, surface pressure, evaporation, solar surface radiation, total column rainwater, large-scale precipitation, and total cloud cover, from the ECMWF (European Centre for Medium-Range Weather Forecasts) Reanalysis v5 (ERA5) dataset, were used as model inputs. Model forecasts were validated against ERA5 rainfall data and compared with the forecasts from the Global Forecast System (GFS) model. Results indicate that the Random Forest model outperforms all others, achieving an RMSE of 6.11 mm and Pearson’s correlation coefficient (R) of 0.74 for a 1-day lead time. The LSTM model achieved an RMSE of 7.46 mm, while the MLP performed less effectively than both RF and LSTM, with an RMSE of 7.61 mm. In comparison, the GFS forecasts displayed an RMSE of 9.16 mm. The RF model outperformed the other models at all lead times; however, its accuracy decreased as the lead time increased. This study highlights the potential of machine learning to improve short to medium range rainfall forecasts, contributing to timely decision-making for urban resilience and resource management.
How to cite: Sarker, M. M., Zobeyer, A. M., Rouf, T., and Rahman, S. M. M.: A Comparative Analysis of Data-Driven Machine Learning Models for Rainfall Forecasting in Bangladesh, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-1959, https://doi.org/10.5194/egusphere-egu25-1959, 2025.
EGU25-2071 | ECS | Posters virtual | VPS5
Enhancing Heavy Rainfall Predictions Over Vulnerable Regions in Assam Using a Spatial Attention-Based Deep Learning NetworkMon, 28 Apr, 14:00–15:45 (CEST) | vP5.14
Forecasting extreme rainfall events (EREs) locally is a major difficulty for meteorological organizations in India's diverse topography, including Assam, Uttarakhand, and Himachal Pradesh. Flash floods cause major socioeconomic damage in certain areas. These extremes are increasingly commonplace during the southwest monsoon season in the country and one of the most destructive EREs occurred in June 2022 and 2023 over Assam. This work explores deep learning (DL) models, specifically spatial attention-based U-Net, in conjunction with simulated daily collected rainfall outputs from different parametrization schemes rainfall output from the Weather Research and Forecasting (WRF) model, considering the limitations of deterministic numerical weather models in accurately forecasting these events. The model trained over the districts of Assam for all days (days 1-4) except the districts where the EREs occurred. The suggested model exhibited a greater ability to predict rainfall at the district scale with a mean absolute error of less than 10 mm over four days in June 2022, outperforming both individual and ensemble outputs of WRF. Furthermore, the suggested model had a high prediction accuracy of 91.9% in categorical rainfall prediction, outperforming WRF models by 51.3%. Furthermore, by accurately forecasting EREs at the district level, including Barpeta, Kamrup, Kokrajhar, and Nalbari, the suggested model has shown improved spatial variation when compared to the WRF model. The suggested DL model is tested for real-time ERE events over Assam in June 2023. In the second part, the model has trained for ERE occurred in 2022 and tested for 2023 over Assam at the district level. The district-level performance of the DL and WRF models is compared, and the DL model performs better than the WRF model in capturing EREs, with a noteworthy accuracy of 54.4% compared to only 22.8% for the WRF model. Notably, the DL model accurately represents the amount and severity of rainfall in Assam's western and southern regions. In summary, the study's conclusions directly affect the development of effective strategies for increased preparedness, mitigation, and adaptation measures over complex hilly regions to lessen the loss of life and property, as well as the improvement of early warning systems and related follow-up action.
How to cite: Trivedi, D., Pattnaik, S., and Sharma, O.: Enhancing Heavy Rainfall Predictions Over Vulnerable Regions in Assam Using a Spatial Attention-Based Deep Learning Network, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-2071, https://doi.org/10.5194/egusphere-egu25-2071, 2025.
