Physical drivers and processes
-Role of the atmosphere, ocean, land, and ice processes in extended-range/S2S predictability;
-Modes of variability (e.g., Madden Julian Oscillation (MJO), quasi-biennial oscillation (QBO), polar vortex strength, and others) impacting the extended-range/S2S predictability;
-Impact of global warming on early warning systems, changes in risks.
Prediction systems
-Evaluation and improvement of S2S prediction systems, including advancements in model physics and comparison between dynamical and data-driven prediction models, data assimilation, ensemble forecasting, and initialization techniques;
-Use of AI/ML methods for S2S prediction, data-driven models, post-processing, and attribution, including innovative techniques for improving forecast accuracy.
Extreme events and early warnings
-Early warnings for single- and multi-hazard events;
-Sources of predictability for extreme events, including multi-hazards events, on the S2S timescale (including driver identification and teleconnections);
-Case studies of extreme or high-impact event prediction and impacts on early warnings;
-Predictability and predictive skill of atmospheric or surface variables, and other variables relevant for socio-economic sectors, such as sea ice, snow cover, soil moisture, and land surface.
Applications and societal relevance
-Sector-specific applications, impact studies on the S2S/extended range timescale;
-Integration of S2S predictions into decision support systems at local, regional, or global levels and co-production of knowledge with stake-holders and decision-makers.
Forecast skill on subseasonal-to-seasonal timescales varies strongly with the large-scale atmospheric state, creating intermittent “windows of opportunity” for skillful prediction. Here we evaluate state-dependent predictability of 2m temperature in subseasonal hindcasts with CESM and in a perfect modelling framework. Skill is quantified as a function of the Pacific-North American pattern, the phase of the El Nino Southern Oscillation, the Madden-Julian Oscillation, the North Atlantic Oscillation and the soil state. Both models exhibit regionally significantly enhanced subseasonal skill during dynamically organized flow regimes like the PNA, El Niño or La Nina, and certain MJO phases, when tropical forcing projects onto an amplified winter jet and supports coherent Rossby wave propagation. The resulting predictability is modulated by the seasonality of the background flow. Our findings demonstrate that regional S2S forecast skill may be higher than suggested by spatial averages. It is investigated if AI generated forecasts can capture this state-dependent predictability.
How to cite: Berner, J., Jaye, A., Richter, J., Mayer, K., and Fowler, M.: On the state-dependent predictability horizon in dynamical and AI forecasts, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21897, https://doi.org/10.5194/egusphere-egu26-21897, 2026.
A set of relaxation experiments with the European Centre for Medium-range Weather Forecasts (ECMWF) model is used to explore the influence of tropical and stratospheric teleconnections on forecast skill, variability of forecast ensemble mean (EM) and ensemble spread (ES) in the wintertime Northern Hemisphere at sub-seasonal timescales. The influence is diagnosed by comparing the relaxation experiments, which relax the temperature and wind fields in specific regions to observed values, with the free running (control) experiment. During weeks 3–6 the tropical relaxation increases the forecast skill for sea level pressure (SLP) mostly south of 50°N but also over the North Atlantic, Northern Europe and eastern Canada. Skill improvements occur via both stratospheric and tropospheric pathways. The stratospheric relaxation improves the skill mostly in high latitudes, over Europe, and North Atlantic. Skill improvements are smaller for surface temperature and total precipitation, suggesting a smaller role of the teleconnections in their predictability.
The increases in skill are generally associated with increased variability of EM, considered to represent the predictable signal, and reduced ES representing noise. However, this does not happen in all areas where the skill is increased. In high latitudes, where the stratospheric impacts are strongest, the EM variability does not increase in the stratospheric relaxation experiments consistently with increases in skill, implying that EM does not reflect well the predictable signal. We suggest that the ensemble size available in the experiments (11 members) is not always enough to make it possible to fully extract signal from noise, and that larger ensembles (20–50 members or even more depending on area and variable) would be beneficial for studies of sub-seasonal predictability associated with the teleconnections in mid- and high latitudes, including windows for forecast opportunities.
How to cite: Karpechko, A., Butler, A., and Vitart, F.: Signal, noise and skill in sub-seasonal forecasts: the role of tropical teleconnections and stratosphere-troposphere coupling, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11008, https://doi.org/10.5194/egusphere-egu26-11008, 2026.
Midlatitude synoptic storms are a major source of forecast uncertainty, yet the mechanisms linking storm track variability, ensemble spread and extreme event predictability remain insufficiently understood. Differences in storm track intensity between the Pacific and the Atlantic sectors span multiple timescales, from daily to seasonal and decadal, and have been linked to basin characteristics, strength of the subtropical jet, the latitudinal position of the jet stream, and to the changes in the cyclone life-cycles. Yet, the relevance of these dynamical processes for forecast uncertainty remains unclear.
This study investigates how the occurrence and intensity of synoptic storms modulate ensemble forecast spread related to the storm track in the North Pacific and North Atlantic, with the goal of identifying sources of forecast uncertainty for midlatitude weather at subseasonal lead times (2–6 weeks lead time). We use ECMWF reforecasts from the Subseasonal to Seasonal (S2S) Prediction Project database, verified against EAR5 reanalysis, to investigate ensemble forecast distributions of the upper-troposphere westerly jet and various storm activity metrics based on eddy kinetic energy (EKE). Our analysis involves a systematic assessment of spread-mean relationships for EKE and other dynamical variables for different midlatitude regions and lead times.
A key result is a robust linear, positive relationship between ensemble mean and spread of EKE in the Atlantic sector, suggesting the contribution of synoptic-scale storms for reliable forecasts. Furthermore, this spread-mean coupling implies that periods of enhanced storm activity, such as wintertime storminess over the North Atlantic, are associated with systematically larger EKE ensemble spread. In contrast, in the Pacific the relationship seems inherently different, indicating a non-trivial role of synoptic storms in forming forecast uncertainty. Large-scale variation of the storm track - such as those associated with teleconnections - are expected to modulate ensemble spread and thereby induce flow-dependent variations in predictability.
The findings highlight the relevance of storm track diagnostics and EKE-based spread metrics as promising tools to improve forecast accuracy and enhance early warning capabilities for high-impact midlatitude storms.
How to cite: Gerstman, H., Rupp, P., Wu, R. W.-Y., Spaeth, J., Vitart, F., and Romppainen-Martius, O.: Linking storm track activity and subseasonal forecast uncertainty in the North Pacific and North Atlantic, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17748, https://doi.org/10.5194/egusphere-egu26-17748, 2026.
Recent advances in machine learning (ML) illustrate the potential for significant improvements in weather predictive skill. At the same time, ML-based technologies have broadened the range of organisations capable of delivering skilful atmospheric forecasts. Given these developments, the ECMWF AI Weather Quest was designed as an open and transparent international competition, enabling a range of organisations to submit ML-based subseasonal forecasts. With participation from more than 40 competing teams spanning academia, public institutions and private companies, the Quest provides a unique framework for systematically evaluating and comparing multiple ML-based subseasonal forecasting systems.
Participants of the AI Weather Quest submit global probabilistic quintile forecasts of near-surface temperature, mean sea level pressure, and precipitation at either a 3- or 4-week lead time in an operational-style forecasting environment. This set-up has encouraged model development whilst challenging participants to develop forecasting systems that operate in realistic settings and deliver actionable forecast parameters.
In this presentation we will provide an overview of the AI Weather Quest design and compare subseasonal forecast skill of both ML-based and dynamical prediction systems. Additionally, we will highlight emerging approaches in ML-based post-processing and fully data-driven forecasting. During the first three-month competitive season, ML-based post-processing of dynamical forecasts achieved the highest skill, highlighting contrasting performance across approaches and underscoring the need for further development in both dynamical and ML-based forecasting. We will also share opportunities for wider engagement and discuss future developments planned for the ECMWF AI Weather Quest.
How to cite: Talib, J., Loegel, O., Vitart, F., Hoffmann, J., and Chantry, M.: Insights from the AI Weather Quest: An international machine-learning competition for sub-seasonal prediction, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3570, https://doi.org/10.5194/egusphere-egu26-3570, 2026.
Subseasonal weather prediction (2–6 weeks lead time) represents a critical "predictability desert" where the influence of atmospheric initial conditions diminishes and boundary forcings have not yet become dominant. Despite its inherent difficulty, skillful subseasonal forecasting is vital for decision-making in agriculture, water resource management, public health and disaster preparedness. While recent breakthroughs in artificial intelligence (AI) and machine learning (ML) have revolutionized synoptic-scale weather forecasting, these gains have not yet fully translated to the subseasonal regime, where systematic biases in both dynamical and data-driven models remain a primary bottleneck.
In this work, we present a novel Probabilistic Bias Correction (PBC) framework that leverages ML to systematically identify and correct errors in global model forecasts. Our approach is model-agnostic, as we demonstrate by showing it can enhance both traditional physics-based dynamical ensembles and emerging AI-based forecasting systems. By training on historical reanalysis and model forecast datasets, the PBC framework significantly reduces systematic errors that typically degrade raw model skill at subseasonal lead times.
We evaluate the performance of our PBC algorithms against several high-standard benchmarks, including climatology, multi-model super-ensembles from major operational centers, and state-of-the-art AI models. Notably, our framework was benchmarked within the context of the AI Weather Quest (sponsored by ECMWF). Results demonstrate that our PBC forecasts outperform all participating dynamical and ML models, including the ECMWF Integrated Forecasting System (IFS) and Artificial Intelligence Integrated Forecasting System (AIFS), in predicting 2-meter temperature, precipitation and mean sea level pressure.
To demonstrate the real-world utility of this system for early warning capabilities, we present case studies of extreme winter weather events in the Eastern United States and Europe. Our model successfully predicted these high-impact events several weeks in advance, with forecasts disseminated in real-time to stakeholders via social media. Our findings suggest that while AI-based models like FuXi-S2S offer a strong alternative to dynamical systems, the integration of probabilistic post-processing is critical to maximize predictive skill and provide reliable, sector-specific decision support in a changing climate.
How to cite: Mouatadid, S., Weyn, J., Guan, H., Orenstein, P., Cohen, J., Mackey, L., Lu, A., Flaspohler, G., Ni, Z., and Dong, H.: Bridging the "Predictability Desert": A Probabilistic Bias Correction Framework for AI and Dynamical Subseasonal Forecasts, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-14077, https://doi.org/10.5194/egusphere-egu26-14077, 2026.
For more than half a century, meteorology has accepted a fundamental limit: weather cannot be predicted beyond two weeks. This boundary, rooted in the chaotic nature of the atmosphere, has shaped generations of forecasting science, defining the boundary between weather and climate forecasting and constraining our ability to anticipate high-impact extremes. At the same time, extreme heat has emerged as the deadliest climate-related hazard worldwide, underscoring the urgent need for reliable early warnings at lead times relevant for public health, energy systems, and disaster risk reduction. Recent advances in deep learning have produced a new class of global weather models accelerating progress in forecasting, raising the question of whether the traditional two-week limit is beginning to shift.
Here we evaluate six state-of-the-art deep learning weather emulators — Pangu-Weather, FuXi, ArchesWeather, AIFS, GraphCast and Aurora — alongside leading dynamical approaches and statistical baselines in forecasting global surface temperature and extreme heat events at a 14-day lead time. Models are evaluated using a suite of metrics, considering global temperature and extreme heat forecasting in both regression and classification settings. Several emulators rival or even surpass physics-based forecasts for temperature, but struggle to balance deterministic skill with realistic spectral properties. While all models display predictive skill for extreme heat, their predictions are deterministic and often inaccurate, offering little insight into uncertainty and limiting their reliability. Overall, results demonstrate that deep learning is starting to extend the frontiers of deterministic predictability. However, key limitations remain that constrain their applicability for operational early-warning systems, highlighting the need for reliable probabilistic approaches in a rapidly warming climate.
How to cite: Decancq, C., Mortier, T., Keune, J., and Miralles, D.: Weather Emulators Push the Frontier of Heat Extremes Forecasting, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6412, https://doi.org/10.5194/egusphere-egu26-6412, 2026.
Forecasting extratropical weather on subseasonal timescales continues to be a challenge. One possible source of atmospheric predictability on these timescales are slowly evolving components of the climate system, most notably tropical modes of variability such as the Madden–Julian Oscillation (MJO) and tropical waves. These sources of predictability are not fully exploited because of systematic errors in numerical weather prediction (NWP) models. In particular, forecast errors that develop in the tropics at lead times of several days grow up-scale, propagate and degrade forecast skill in the extratropics on subseasonal timescales. Regions of forecast errors that exert the strongest influence on extratropical forecast skill remain poorly identified. Relaxation experiments using NWP models provide a means to isolate these regions, but such experiments are computationally demanding. In this study, we employ machine learning–based weather prediction models to perform relaxation experiments across multiple tropical regions. Probabilistic forecasts are generated using perturbed initial conditions from the Ensemble of Data Assimilations (EDA) of the European Centre for Medium-Range Weather Forecasts (ECMWF). Reforecasts covering a five-year period (2020--2024) are used to systematically assess the impact of relaxation strategies and the role of tropical variability, with a particular focus on the MJO, on extratropical subseasonal forecast skill. A key-finding is that predictions of the negative phase of the North Atlantic Oscillation improve when relaxation is applied following MJO phases 6 and 7 at initial time. Rossby wave source diagnostics are examined to investigate the dynamical processes leading to improvements in extratropical forecasts. The results demonstrate the value of relaxation experiment as a diagnostic tool when integrated with emerging machine learning–based prediction systems.
How to cite: Li, S., Namdev, P., and Quinting, J.: Impact of Systematic Relaxation Experiments on Subseasonal Forecast Skill in Machine Learning Weather Models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4190, https://doi.org/10.5194/egusphere-egu26-4190, 2026.
Heatwaves are among the most impactful climate-related hazards in Europe, with increasing frequency, duration, and severity under climate change. Currently, heatwave warnings across Europe are typically issued 2–7 days in advance. Extending these lead times could substantially enhance preparedness by enabling earlier adaptive actions and more effective resource allocation.
On sub-seasonal time scales, extended-range weather forecasts (approximately 2 weeks to 1 month) have been shown to exhibit higher skill for warm extremes than for average temperature conditions over Europe. In particular, the persistence of prolonged heat waves seems to have a higher-than-average level of predictability even at a 3-week lead time. Recent verification studies have demonstrated statistically significant probabilistic skill of ECMWF extended-range ensemble reforecasts for European heatwaves defined using 5-day mean temperatures. Together, these findings indicate that extended-range ensemble forecasts can provide early warning information not only on the likelihood of heatwave occurrence, but also on the potential persistence and severity of extreme heat events.
Building on this demonstrated predictability, we take the next step towards practical early warning applications by developing a forecast and alert logic for extended-range heatwave prediction. We post-process ECMWF extended-range ensemble forecasts to produce probabilistic heatwave forecasts using a previously developed and verified methodology based on 5-day mean temperature thresholds. Building on this framework, we develop a rule-based alert logic that translates probabilistic forecasts into actionable early warning information. The alert logic combines forecasted heatwave probabilities with indicators of forecast reliability and flow-dependent predictability, including ensemble spread, the heatwave life cycle state at forecast initialization, and North Atlantic sea surface temperature anomalies at the time of forecast initialization. These components are used to define multiple alert levels, ranging from no-action conditions to high-impact heatwave risk.
The proposed framework provides a practical pathway from extended-range probabilistic forecasts to impact-oriented early warning.
How to cite: Korhonen, N. and Gregow, H.: Development of Forecast and Alert Logic for Extended-Range Heatwave Early Warning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5602, https://doi.org/10.5194/egusphere-egu26-5602, 2026.
Intraseasonal variation of winter climate in China has been remarkable in recent years, such as: reversed or alternating extreme cold and extreme warm events in different months or in different stages of the winter. There are many challenges in climate prediction in winter because the intraseasonal climate variation is often within the seasonal mean variation. It is therefore urgent to understand the intraseasonal variation of winter climate in China, to identify its predictability and predictive sources, and to propose effective prediction methods and prediction models for it. The author reviews progress in research during the last five years on the main characteristics, physical processes, mechanisms, predictability, and prediction of intraseasonal variation of winter climate in China, considering several related systems including the winter monsoon, Siberian high, and stratospheric polar vortex.
How to cite: Fan, K.: Intraseasonal variation of winter climate in China and climate prediction, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1041, https://doi.org/10.5194/egusphere-egu26-1041, 2026.
The subseasonal-to-seasonal (S2S) prediction gap is a major challenge in operational forecasting, especially for the Indian Summer Monsoon. Prediction skill of the dynamical models for its dominant mode of variability, the Monsoon Intraseasonal Oscillation (MISO), drops sharply beyond one week. This 30–60-day northward-propagating mode governs active and break spells, with major implications for agriculture, water resources, and disaster preparedness across South Asia. In this study, we developed a deep learning forecasting framework built on a Transformer architecture to capture the long-range dependencies inherent to intraseasonal variability. We used 25 years of high-resolution (0.25° × 0.25°) daily TRMM/GPM precipitation data to derive MISO indices (MISO1 and MISO2) via extended empirical orthogonal function (EEOF) analysis for the boreal summer months (June–September). These indices formed the basis for training and evaluating the Transformer model. When evaluated for the 2018–2022 period, the Transformer substantially outperformed traditional numerical weather prediction models, accurately forecasting the phase and amplitude of MISO with lead times of up to 18 days. It also produced better phase alignment and reduced phase-lag errors compared to NWP systems at extended leads. The approach was further extended to predict NLSA-based MISO indices. In addition, a Vision Transformer (ViT) was used to make preliminary forecasts of spatial rainfall patterns associated with MISO propagation. These results highlight the potential of advanced deep learning architectures to enhance S2S prediction of monsoon intraseasonal variability, supporting improved early warning systems and decision-making in monsoon-affected regions.
How to cite: Keloth Methal, A., Raj, P., Sukumaran, S., and Kodamana, H.: Capturing Long-Range Dependencies for Improved MISO Prediction via Deep Learning , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1176, https://doi.org/10.5194/egusphere-egu26-1176, 2026.
Improving subseasonal predictions of heatwave (HW) onset is crucial for early warning systems. While soil moisture (SM) is recognized as a key initial land surface condition, the impact of its three-dimensional (3D) error structure on HW onset prediction uncertainty, and strategies to mitigate this uncertainty, remain insufficiently explored.
This two-part study addresses these gaps, focusing on predictions with a three-week lead time for eight HW onsets over the Middle and Lower Reaches of the Yangtze River (MLYR) region. First, the Conditional Nonlinear Optimal Perturbation (CNOP) method was employed to identify the 3D-structured initial SM errors that maximize uncertainty in subseasonal HW onset predictions. Results show that these structured CNOP-type errors, characterized primarily by negative anomalies with coherent vertical patterns, intensify HW magnitude and advance onset timing. They exert greater impact than spatial random errors by altering surface energy partitioning: reducing latent heat and enhancing sensible heat primarily through vegetation-related processes, while also modulating net longwave radiation via the Stefan-Boltzmann law. Further experiments revealed the importance of deep-layer SM errors and nonlinear synergistic effects across soil layers.
Building on this, the second part evaluates whether targeted observations of initial SM in CNOP-identified sensitive areas (SAs) can enhance prediction skill. Observing System Simulation Experiments (OSSEs) for eight HW events demonstrate that initializing with more realistic SM over SAs consistently outperforms improvements over non-sensitive areas. This targeted approach improves predictions for an average of 86% of ensemble members per case and reduces the mean error in area-averaged maximum temperature during HW onset by 43%. The improvement is attributed to more accurate initial SM conditions, leading to a better representation of surface heat fluxes.
Collectively, these studies systematically highlight the error structure of initial SM field as a key source of subseasonal HW prediction uncertainty and demonstrate the practical potential of CNOP-based targeted observation strategies to improve HW onset predictions.
How to cite: Liu, H., Sun, G., Mu, M., Zhang, Q., and Chen, B.: From Error Identification to Targeted Observations: The Role of 3D Soil Moisture Errors in Improving Subseasonal Heatwave Onset Predictions, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2229, https://doi.org/10.5194/egusphere-egu26-2229, 2026.
Unlike traditional droughts that evolve gradually, flash droughts (FD) are characterized by rapid intensification, leading to sustained dry conditions with disproportionately high impacts on ecosystems. Despite substantial progress in short- to medium-range weather forecasting, predicting these events remains a significant hurdle for both early warning systems and physically-based subseasonal-to-seasonal (S2S) prediction frameworks.
To address this challenge, we present a deep learning framework leveraging a Vision Transformer with explicit temporal attention for the prediction of soil moisture anomalies (SMA) over Europe. The model employs a dual-stream attention mechanism that disentangles temporal dynamics from spatial dependencies: temporal self-attention with rotary positional embeddings captures lead-time-dependent evolution at each location, while spatial attention encodes cross-regional relationships. This architecture enables to learn multi-scale representations, ranging from synoptic variability to persistent anomaly patterns. Furthermore, the model supports probabilistic forecasting, estimating the full conditional distribution of soil moisture anomalies to provide principled uncertainty quantification, a critical requirement for operational early warning systems.
Additionally, the framework employs a multitask learning approach that exploits the relationship between continuous soil moisture anomalies and discrete flash drought characteristics derived from the Flash Drought Intensity Index (FDII). This index integrates both the rate of soil moisture decline and drought severity into a unified indicator. A shared encoder learns representations that capture the coupled dynamics of soil moisture evolution and flash drought emergence, while task-specific prediction heads accommodate the distinct statistical properties of each target variable.
The results indicate that our approach achieves predictive skill competitive with more complex spatio-temporal models while maintaining computational efficiency suitable for operational deployment. Evaluation against standard baselines, including climatology and persistence, as well as state-of-the-art deep learning models, demonstrates the framework’s ability to resolve the rapid intensification dynamics typical of flash drought onset. This work lays the foundation for interpretable, scalable, and probabilistic prediction of rapid-onset drought events at S2S timescales.
How to cite: Otero, N., Fernández-Torres, M. Á., Özer, A., and Ma, J.: Towards Sub-Seasonal Flash Drought Prediction Using a Vision Transformer Approach, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3518, https://doi.org/10.5194/egusphere-egu26-3518, 2026.
A multi-scale coupled framework has been implemented in the Central Weather Administration’s Global Ensemble Prediction System Version 3 (CWA GEPSv3) to improve extended-range forecasts over Taiwan. Reforecasts for January from 2001 to 2020 show skillful Madden–Julian Oscillation (MJO) predictions, with an average lead time of 17 days and a maximum of 33 days. The model realistically captures the eastward propagation of the MJO from the Indian Ocean to the Maritime Continent (MC) but fails to sustain its intensity beyond the MC due to a boundary-layer dry bias emerging 5–10 days before the convection center’s arrival.
Event-based analysis reveals that accurate MJO forecasts are more common during La Niña years, whereas poor forecasts occur more often during El Niño years. The low-frequency moisture field mitigates the dry bias over the MC during La Niñas, but amplifies it during El Niños. Ocean–atmosphere coupling enhances forecast skill at 20–30 lead days, and it is only pronounced for the good-prediction cases.
The boundary-layer dry bias over the MC primarily results from weak upward motion linked to insufficient meridional convergence. A modified dynamical core enhances the simulation of horizontal convergence, yielding clearer eastward propagation of MJO signals. These results elucidate the physical processes underlying model biases in GEPSv3 and provide practical guidance for improving subseasonal-to-seasonal forecasting.
How to cite: Tseng, Y., Lee, C.-W., Shao, Y.-C., Liu, P.-Y., Tseng, H.-H., and Chen, J.-H.: MJO Prediction in the CWA GEPSv3: Model Performance and Physical Processes Underlying Simulation Errors, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4276, https://doi.org/10.5194/egusphere-egu26-4276, 2026.
Data-driven models have achieved significant progress in sub-seasonal prediction, but predicting the initiation of the Madden-Julian Oscillation (MJO) remains a critical challenge, largely due to initial uncertainties from sparse observations over tropical oceans and complex multiscale interactions. Therefore, identifying sensitive areas in initial conditions is crucial to both reveal the underlying error growth mechanisms and provide guidance for target observations. Here, the FuXi-S2S model is applied to explore the initial sensitivity and instability modes of MJO initiation. First, the evaluation of prediction skill identifies the initiation of primary MJO events at a 3-pentad lead time as a critical bottleneck. Simulations initialized with optimized initial conditions within analysis uncertainty closely reproduce the observed MJO evolution, thereby validating the high initial sensitivity during the first 4 pentads. Subsequently, the conditional nonlinear optimal perturbation (CNOP) method is utilized to identify the optimally growing initial errors (OGIEs) and optimal precursors (OPRs). Analysis of OGIEs reveals three dominant types of error modes causing the largest forecast errors, indicating that the rapid growth of OGIEs is driven by the coupling of local low-level thermodynamic instability (temperature and moisture) and upstream upper-level dynamic forcing (wind). Moreover, the spatial structure and perturbation evolution of OPRs exhibit high consistency with OGIEs. The identification of these shared instability modes provides a theoretical foundation for target observations, suggesting that additional observations in sensitive areas can simultaneously reduce initial errors and capture precursors.
How to cite: Peng, Z., Mu, M., and Li, H.: Predictability of MJO Initiation in Data-Driven Models: Shared Instability Modes of Optimally Growing Initial Errors and Optimal Precursors, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4347, https://doi.org/10.5194/egusphere-egu26-4347, 2026.
Sudden stratospheric warming (SSW) is identified as key sources of skill in winter subseasonal-to-seasonal (S2S) forecasts because of their surface impacts lasting up to 30–60 days through stratosphere troposphere coupling, despite their typical prediction being limited about 10 days. A better understanding of the predictability of the SSW itself, thus, is fundamental. Most of the previous studies investigate the predictability of SSW events using linear approaches, which are insufficient given the inherently chaotic and nonlinear nature of SSWs. In the study, we apply a nonlinear method—Backward Searching for the Initial Condition (BaSIC)—to quantify the local predictability limit the 2021 SSW event, which caused cold extremes across East Asia and North America. Using ERA5 reanalysis and the S2S reforecasts data, BaSIC estimates the maximum prediction lead time of this 2021 SSW event to be 17 days. To explore sensitive region of forecast uncertainty, we identify regions of fastest error growth via error tracking in S2S systems using BaSIC method. Forecast errors during the SSW event are small across the polar stratosphere after initiation but grow gradually over two weeks, accelerating rapidly over central Eurasia (30◦-60◦E) and spreading across the continent. This points to central Eurasia at high altitudes as a critical region for SSW forecast error development.
How to cite: li, X.: Quantifying the practical local predictability of the 2021 sudden stratospheric warming event using a nonlinear method, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5219, https://doi.org/10.5194/egusphere-egu26-5219, 2026.
Intraseasonal variations of sea surface temperature (SST) in the South China Sea (SCS) during winter are investigated by using atmospheric and oceanic reanalysis in this study. The dominant pattern of SST variations within the 10–90-day timescale is derived by empirical orthogonal function analysis, which features a basin-wide warming or cooling spatial pattern, with a cycle period of 30–35 days. Composite analysis and mixed-layer heat budget analysis are conducted to investigate the physical process controlling SST variability. The formation of intraseasonal SST variations is primarily attributable to the forcing of wind-related latent heat flux and shortwave radiation flux changes. During the warming (cooling) period, anomalous southerly (northerly) winds tend to weaken (enhance) climatological northerly winds. This, in turn, results in a weakening (enhancement) of wind speed, favoring a reduction (increase) in latent heat flux from the ocean into atmosphere, accompanied by an increase (decrease) in shortwave radiation flux. In addition to surface heat flux forcings, ocean zonal advection is the second most significant contributing factor, exerting a negative effect. Finally, the effect of the Madden–Julian Oscillation (MJO) on the SST is studied. The variations in wind anomalies and surface heat flux changes associated with SST intraseasonal variability are significantly related to the MJO activities. The anomalous anticyclone (cyclone) in the northwest Pacific Ocean is induced by MJO, with enhanced (depressed) convection occurring in the Indian Ocean and depressed (enhanced) convection over the Maritime Continent and western Pacific Ocean, which accounts for the anomalous southerly (northerly) winds and enhanced (depressed) shortwave radiation observed.
How to cite: Mingyue, Q., Ren, X., Wu, X., and Wang, G.: Intraseasonal SST variations in the South China Sea during the borealwinter and the impacts of Madden-Julian Oscillation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6531, https://doi.org/10.5194/egusphere-egu26-6531, 2026.
Based on the hindcasts from five subseasonal-to-seasonal (S2S) models participating in the S2S Prediction Project,
this study evaluates the performance of the multimodel ensemble (MME) approach in predicting the subseasonal
precipitation anomalies during summer in China and reveals the contributions of possible driving factors. The results
suggest that while single-model ensembles (SMEs) exhibit constrained predictive skills within a limited forecast lead time
of three pentads, the MME illustrates an enhanced predictive skill at a lead time of up to four pentads, and even six pentads,
in southern China. Based on both deterministic and probabilistic verification metrics, the MME consistently outperforms
SMEs, with a more evident advantage observed in probabilistic forecasting. The superior performance of the MME is
primarily attributable to the increase in ensemble size, and the enhanced model diversity is also a contributing factor. The
reliability of probabilistic skill is largely improved due to the increase in ensemble members, while the resolution term does
not exhibit consistent improvement. Furthermore, the Madden–Julian Oscillation (MJO) is revealed as the primary driving
factor for the successful prediction of summer precipitation in China using the MME. The improvement by the MME is not
solely attributable to the enhancement in the inherent predictive capacity of the MJO itself, but derives from its capability in
capturing the more realistic relationship between the MJO and subseasonal precipitation anomalies in China. This study
establishes a scientific foundation for acknowledging the advantageous predictive capability of the MME approach in
subseasonal predictions of summer precipitation in China, and sheds light on further improving S2S predictions.
How to cite: Guo, L.: Advantages of the Multimodel Ensemble Approach forSubseasonal Precipitation Prediction in Chinaand the Driving Factor of the MJO, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9223, https://doi.org/10.5194/egusphere-egu26-9223, 2026.
The Madden-Julian Oscillation (MJO) is the dominant and most influential intraseasonal oscillation in the tropics, as well as a prominent source of subseasonal-to-seasonal predictability. Its maintenance and propagation, particularly across the Maritime Continent, remain a challenge for theory, simulation and forecasting. In modern MJO theories, moisture plays a central role. As the primary source of moisture, sea surface latent heat flux has received significant attention. The sea surface latent heat flux can be induced by both base-state winds and wind bursts, but the different effects of these two types of latent heat flux are not well understood.
This study addresses this issue by analyzing CESM hindcasts of a strong MJO event in April 2009 that crossed the Maritime Continent with little decay. The control simulation reproduces the observed propagation characteristics and intensity of this event. We then impose an upper limit on wind speed in the bulk formula used by the model to calculate sea surface latent heat flux. By lowering this limit in different simulations, we reduce the latent heat flux from strong winds first, followed by the flux from base-state winds. Based on frequency distributions over the Indian Ocean and the Maritime Continent, the peaks for base-state winds and wind bursts occur at 3 m/s and 10 m/s respectively.
Results indicate that wind-burst-induced latent heat flux is essential to maintaining MJO amplitude over the Maritime Continent, due to the region’s lower sea fraction. As convection over land tends to disrupt the coherent organization of the MJO convection envelope, lower sea fraction increases the sensitive of MJO amplitude to sea surface latent heat flux. On the other hand, the base-state surface latent heat flux modulates MJO propagation speed due to its effectiveness in moistening the atmosphere. As the base-state latent heat flux is reduced, the atmosphere dries, moisture advection decreases, and the MJO slows down. Additional simulations confirm these findings in other MJO cases. This study underscores the importance of accurately simulating strong winds for maintaining MJO amplitude over the Maritime Continent and overcoming the barrier effect.
How to cite: Weng, L. and Lin, Y.: Distinct Roles of Base-State and Wind-Burst-Induced Sea Surface Latent Heat Flux in MJO Maintenance and Propagation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9451, https://doi.org/10.5194/egusphere-egu26-9451, 2026.
Energy demand, especially for residential heating, is largely driven by temperature. High-quality subseasonal-to-seasonsl (S2S) forecasts of temperature are therefore valuable for risk management and energy trading, yet the use of these forecasts is often limited by their complexity, by difficulties in combining different forecast types, and by the relatively weak probabilistic skill they produce. Sequential learning algorithms (SLA), offer a means to overcome many of these difficulties. SLAs optimally combine information from multiple ‘experts’ or predictors using weights and reduce forecast bias by continuously learning over time from each forecast verification. The information from these ‘experts’, which can both be statistical or numerical forecasts, are used as SLA inputs and blend into a single information stream through dynamical updating of the weights. Here, ‘experts’ are defined as quantiles of raw ECMWF S2S 2 m temperatures (T2m) forecasts without bias adjustment and ERA5 T2m climatology. The SLA produces probabilistic forecasts of Great Britain-averaged T2m at lead times of 1-4 weeks for the period 2004-2023. Results show positive anomaly correlation co-efficient and rank probability skill scores for SLA T2m forecasts across all weeks compared to both the raw S2S and climatological forecast. Analysis of the weight evolution shows that SLA relies heavily on the raw forecast experts at weeks 1-2 but shifts towards climatological experts in the later weeks, with a clear seasonal evolution to the weight profile. It is also confirmed that this online-learning approach with adaptive weights outperforms the most optimal static weight combination even though the latter is permitted the benefit of perfect foresight.
How to cite: Das, A., Brayshaw, D., Methven, J., Frame, T., O'Reilly, C., Sharma, S., Mammatt, J., and Fox, S.: Using machine learning to enhance skill of subseasonal-to-seasonal (S2S) temperature forecasts, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11985, https://doi.org/10.5194/egusphere-egu26-11985, 2026.
Organized mesoscale convective systems (MCSs) over the Arabian Peninsula (AP) are a major driver of extreme precipitation and flash flooding in the cool season (October - April). Existing criteria for MCS tracking methods do not capture this phenomenon over the AP, including several record-breaking MCS-driven extreme precipitation events that caused significant socioeconomic losses in the Kingdom of Saudi Arabia (KSA). In this study, we evaluate the MCS tracking capability and calibrate regional tracking criteria for the AP.
Based on several MCS-driven precipitation events over the past 20 years, AP MCS criteria are updated as follows: size over 20,000 km2, cloud-top temperature less than 230 K, and merging/splitting duration over 3 hours. The AP MCS tracking criteria are also updated specifically for application to convective-permitting Weather Research and Forecasting model (WRF) output. WRF MCS size and durations are similar to observed MCSs, but the core cloud temperature threshold is lowered to 218 K.
The MCS tracking algorithm is then applied to a 20-year MERGIR brightness temperature (Tb) dataset and a corresponding 20-year subseasonal WRF (4-km grid spacing) ensemble reforecast product. The WRF subseasonal ensemble reforecasts are available at 1-week to 4-week lead times. Forecast skill is assessed using categorical statistics such as the Critical Success Index, combined with a neighborhood verification method to reduce double-penalty effects.
The AP MCS tracking results based on WRF subseasonal ensembles exhibit robust tracking capability in both early and late cool season, with respect to seasonal climatology and extreme convective case studies. The convective-permitting reforecast demonstrates subseasonal forecast skill and the potential to enhance early warning capabilities for public safety and disaster risk mitigation.
How to cite: Chang, H.-I., Risanto, C. B., Castro, C. L., Luong, T., and Hoteit, I.: Developing tracking capability for mesoscale convective systems in the Arabian Peninsula through observations and model-based subseasonal reforecasts, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-14647, https://doi.org/10.5194/egusphere-egu26-14647, 2026.
False monsoon onsets involve an early-season wet spell followed by a prolonged dry spell, often resulting in agricultural losses when sowing is initiated during premature rains and farmers are unprepared for the dry conditions. Despite its importance for risk reduction for hundreds of millions of farmers in the tropics, the predictability of these pre-monsoonal wet–to–dry events remains largely unexplored. Here, we benchmark six state-of-the-art artificial intelligence weather prediction (AIWP) models (AIFS, FuXi, FuXi-S2S, GraphCast, GenCast, NeuralGCM) and a numerical weather prediction (NWP) model (IFS) against novel, decision-relevant historical reference forecasts to assess the ability to predict the false monsoon onset at lead times up to 30 days. We find that both AIWP and NWP models exhibit positive predictive skill in the core monsoon zone of India, with ensemble-based probabilistic models retaining positive predictive value relative to these reference forecasts across all lead times. Deterministic skills vary strongly with regions, with good short-lead predictability (0-10 days) and a decrease in skills at longer lead times (11-30 days). We further evaluated the models using well-documented canonical false onset events from the literature and found that skillful forecasts are associated with the ability to reproduce the large-scale circulation evolution characteristic of false onsets, in particular the progression from a transient monsoon-like state to a subsequent circulation collapse that produces a dry spell.
We use agriculturally relevant thresholds to define monsoon onset, wet spells, and dry spells. To enable a meaningful assessment of model skill, the reference forecast is constructed from 124 years of gridded rain-gauge observations and quantifies the baseline probability of false monsoon onsets within a decision-relevant framework. We first calibrate model-specific event-definition wet- and dry-spell thresholds using quantile mapping within a leave-one-year-out cross-validation framework, rather than applying bias correction directly to rainfall fields. Forecast performance is evaluated using deterministic and probabilistic metrics, including probability of detection, false alarm ratio, critical success index, and Brier score. Reliability diagrams show systematic overconfidence at higher forecast probabilities, indicating the need for additional calibration and post-processing. Together, this framework establishes a decision-relevant benchmark and evaluates current AI-based and physics-based forecast systems for the sub-seasonal early warning of false onsets involving dry spells.
How to cite: Gupta, M., Aitken, C., Masiwal, R., Marchakitus, A., Boos, W., Kowal, K., Jina, A., and Hassanzadeh, P.: Operational benchmarking of AI and NWP models for false monsoon onset prediction in India, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17228, https://doi.org/10.5194/egusphere-egu26-17228, 2026.
The Blue Nile Basin, located in the Greater Horn of Africa, is highly vulnerable to climate variability, where hydrometeorological extremes have severe socio-economic consequences. Reliable prediction on sub-seasonal to seasonal (S2S) timescales is therefore critical for preparedness in sectors such as agriculture, water and dam management. However, S2S prediction in the region remains particularly challenging due to complex orography, the influence of large water bodies, various climate zones, and the interaction of multiple large-scale circulation modes, including ENSO, the Indian Ocean Dipole, and the Madden–Julian Oscillation.
While global forecasting systems provide valuable large-scale climate information, their direct application at regional scale is limited. In heterogeneous regions such as the Blue Nile Basin, global products are often insufficient in spatial resolution and show systematic biases that reduce their usability for regional and sector-specific applications. This requires targeted post-processing to correct errors and regionally enhance the forecasts.
In this study, we evaluate state-of-the-art seasonal and sub-seasonal forecasting products from the ECMWF, focusing on the SEAS5 seasonal forecasting system (lead times up to 215 days) and the ECMWF sub-seasonal range forecasts (lead times up to 46 days). Forecast skill is assessed against ERA5 and ERA5-Land reanalyses, as well as a composite observational dataset combining satellite and station measurements (CHIRPS). To enhance the raw forecasts, we apply an established statistical post-processing technique, namely bias correction and spatial disaggregation (BCSD), alongside advanced deep learning approaches. The latter include Seasonal AFNO-based models and ProS2St, which has previously been developed and tested at global scale within the ECMWF AI Weather Quest Challenge.
Our results demonstrate that post-processing methods significantly improve raw forecast performance over the Blue Nile Basin. Despite these improvements, outperforming climatology remains challenging for meteorological variables alone. However, we show that when the enhanced forecasts are used as input for subsequent impact models, such as hydrological models, they provide added value compared to climatological forcing.
This work highlights the potential of regionally enhanced meteorological forecasts as a foundation for sub-seasonal to seasonal prediction systems. By coupling post-processed meteorological forecasts with hydrological and crop models, we enable S2S forecasts that support improved decision-making in specific sectors. The focused evaluation of S2S forecasting products and post-processing methods for the Blue Nile Basin, together with their integration into downstream impact models, represents a novel contribution toward operational and application-oriented prediction systems in the region.
How to cite: Wiegels, R., Chwala, C., Polz, J., Glawion, L., Lorenz, C., Weber, J. N., Hageltom, Y., Sawadogo, W., Schober, T. C., Janner, S., Zargar, M., Bronstert, A., and Kunstmann, H.: Toward a Seamless Sub-Seasonal to Seasonal Prediction System for the Blue Nile Basin, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18144, https://doi.org/10.5194/egusphere-egu26-18144, 2026.
Posters virtual: Mon, 4 May, 14:00–18:00 | vPoster spot 5
EGU26-16394 | ECS | Posters virtual | VPS2
S2S Forecast Skill Assessment for Summer Monsoon Drought WarningMon, 04 May, 14:06–14:09 (CEST) vPoster spot 5
Persistent extreme weather anomalies lasting several weeks to months can lead to drought and compound dry–hot extremes, posing serious socio-economic risks in the Indian monsoon region. Although subseasonal-to-seasonal (S2S) prediction systems have advanced, the extent to which these models represent drought-relevant hydroclimatic variability over India has not been adequately quantified. Here, we focus on evaluating the hindcast quality of weekly accumulated precipitation and temperature from multiple S2S models with lead times up to six weeks during the JJAS season over India. These model hindcast outputs and IMD observations are regridded to a common 0.5° resolution and analyzed using deterministic forecast skill metrics at various lead times, statistical bias correction is then applied to isolate systematic model errors, followed by SPI-based drought diagnostics, and compound dry–hot extreme indices are derived and computed. The analysis reveals modest forecast skill at early lead times, followed by a systematic decline as lead time increases, with precipitation predictability deteriorating more rapidly than temperature predictability. Although the models generally capture the large-scale spatial distribution of drought-prone regions, they significantly underestimate the frequency and spatial extent of compound dry–hot conditions, exhibiting pronounced regional dependence across India. These results highlight key limitations and identify opportunities to enhance subseasonal drought early-warning systems.
Keywords: Subseasonal-to-seasonal predictability, Indian Summer Monsoon, Drought, Climate extremes, Hindcast evaluation
How to cite: Sukumaran, S. and Agarwal, A.: S2S Forecast Skill Assessment for Summer Monsoon Drought Warning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16394, https://doi.org/10.5194/egusphere-egu26-16394, 2026.
EGU26-20505 | ECS | Posters virtual | VPS2
Cost-Effective ECMWF AIFS Ensemble Inference for Subseasonal Forecasting in East AfricaMon, 04 May, 14:09–14:12 (CEST) vPoster spot 5
In Eastern Africa, subseasonal forecasts are critical for early warning systems as climate extremes severely impact food security and livelihoods. ECMWF Artificial Intelligence Forecasting System (AIFS ENS v1.0) , an ensemble-based probabilistic data-driven forecast model developed by ECMWF offers unprecedented opportunities for regional applications through AI-driven weather prediction, but GPU compute costs and data access challenges limit deployment. As participants in the ECMWF AI Weather Quest, we developed solutions enabling cost-effective, cloud-based AIFS ensemble forecasting tailored for regional climate centers.
We implemented a workflow (https://github.com/icpac-igad/ea-aifs) leveraging Google Cloud Platform infrastructure. Initial conditions are accessed via ECMWF's IFS data stored at AWS (Amazon Web Service) open data program at S3 Cloud storage using GRIB index-kerchunk, and VirtualiZarr methods for efficient data streaming without local storage overhead. The workflow employs experimental FP16 (half-precision) inference on AIFS ensemble models along with the standard FP32, evaluating GPU memory requirements and enabling deployment on cost-effective T4/L4 GPUs rather than expensive A100 instances.
Verification results from the SON (September-October-November) 2025 season as part of the AI Weather Quest demonstrates that Team Fahamu's submission using AIFS ensemble forecasts for temperature and mean sea-level pressure outperforms climatology benchmarks. Regional evaluation over East Africa reveals promising subseasonal skill for temperature at lead times of 2-4 weeks—critical timescales for agricultural planning and anticipatory drought/flood action—while evaluation of precipitation forecasts is ongoing. This method provides a scalable template for regional climate centers globally to operationalize state-of-the-art AI weather models cost-effectively, advancing the democratization of advanced forecasting capabilities.
How to cite: Koech, E., Kalladath, N., Mwanthi, A., Ogelo, A., Kinyua, J., Koros, H., Lelaono, M., Misiani, H., Bekele, T., Seid, H., Gudoshava, M., and Amdihun, A.: Cost-Effective ECMWF AIFS Ensemble Inference for Subseasonal Forecasting in East Africa , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20505, https://doi.org/10.5194/egusphere-egu26-20505, 2026.
