Annual Seminar 2025

As an introduction to the session “Serving a future society”, we will take look back at ECMWF’s contributions in shaping the future of meteorological and environmental services, working hand in hand with its Member and Co-Operating States and the research community. The presentation will highlight how advances in predictability science and technology have made their way into applications to address societal needs, driving the seamless integration of weather, environmental, and climate data and products.

This talk will highlight some of the key ECMWF research directions in the context of its new strategy.

The ECMWF Vision is "World-leading monitoring and predictions of the Earth System enabled by cutting-edge physical, computational and data science, resulting from a close collaboration between ECMWF and the members of the European Meteorological Infrastructure, will contribute to a safe a thriving society".

As well as emphasizing the ultimate purpose and the importance of collaborations, this wording very deliberately emphasizes all of physical, computational and data science. In terms of physical science, the needs to continue to improve physical models (for use across timescales) and data assimilation systems (e.g. to more fully exploit satellite data over difficult surfaces) remains. Moving to higher resolution is an exciting opportunity, but also needs developments in physical and computational science (e.g. adapting sustainably to evolving computer architectures and handling ever-increasing amounts of data). Finally, the revolution in the application of artificial intelligence in the meteorological field means that physics-based systems will increasingly have a hybrid component, and fully data-driven systems are confidently expected to be a major and possibly dominant contributor to operational systems in the coming years. These systems typically learn from reanalysis - which will remain a major ECMWF focus both for this purpose and for climate monitoring - and hence remain anchored on the physics-based systems.

Extreme weather events are causing unprecedented floods, droughts, fires, and ecosystem damage on all scales. This necessitates better early warnings and climate and weather services, especially for transitioning to carbon-neutral economies. These challenges demand faster innovation, technological advancements, and stronger interdisciplinary collaboration, with a significant role for the private sector. The UN's "Early Warnings for All' initiative aims to establish crucial early warning systems.

Urbanization and our changing global economy have increased the need for accurate projections of climate change and improved seamless numerical weather predictions of disruptive and potentially beneficial weather events at the convective scale. Technological innovations are also leading to an evolving and growing role for the private sector in the weather and climate enterprise. The challenges faced in accelerating advances in weather and climate forecasting propose a vision for key actions needed across the private, public, and academic sectors.

Actions span:

(i) Utilizing new observational and computing ecosystems;
(ii) Developing strategies to advance Earth system models;
(iii) Exploring ways to benefit from the growing role of artificial intelligence;
(iv) Implementing practices to improve the communication of forecast information and decision support in our age of the internet and social media;
(v) Addressing the need to reduce the relatively large and detrimental impacts of weather and climate on all nations, especially low-income nations.

Regarding artificial intelligence, there will be a need to increase investment in data science, as well as a better understanding of predictability, physical and dynamical processes. We will illustrate this in the context of subseasonal-to-seasonal predictions.

The performance of machine learning weather prediction (MLWP) models (e.g., GraphCast, Pangu-Weather) has proved better than numerical weather prediction (NWP) models on the synoptic scale analysis and specific weather events analysis (e.g., typhoon track) at global scale from multiple studies in recent two years. Most studies not only show the comparison of forecast performance between NWP and MLWP models but also specify the limitations in current MLWP models. In particular, most MLWP models provide meteorological variables with the horizontal resolution of 0.25o and without precipitation, which could not have enough information to satisfy the requirements (e.g., disaster prevention) of small or complex terrain regions. Therefore, the method of generating higher horizontal resolution meteorological variables by taking advantage of combining the information both from NWP and MLWP is proposed in this study, which would provide detailed information to make a reliable weather forecast.

This study aims to derive high horizontal resolution quantitative precipitation forecast (QPF) by integration of MLMP typhoon track prediction and Central Weather Administration (CWA) Ensemble Typhoon Quantitative Precipitation Forecast (ETQPF) model which is based on the characteristic of rainfall terrain locking effect. Three initial conditions from physics-based NWP models are used as inputs for six MLWP models to produce the MLWP-based multi-model ensemble typhoon tracks. The ensemble mean track is then applied to produce the QPF based on ETQPF. During super typhoon Gaemi (202403), the MLWP-based ensemble mean track indicates moderate improvement as compared to NWP models. In addition, the integration of MLWP-based ensemble mean typhoon track into ETQPF shows promising results on the QPF. From the study findings, adequate integration of NWP and AI/ML could provide positive impacts on the weather forecast, which drives CWA to enhance the development of NWP, AI/ML, and both integration applications in future.

This presentation will explore how the use of storylines, employing spectral nudging, is advancing our ability to generate detailed, physically plausible narratives of how climate change is impacting weather and environmental events, including extremes. Spectral nudging selectively constrains the evolution of large-scale atmospheric winds (i.e., the jet stream) in global coupled climate simulations to follow observed trajectories while allowing finer, localized details, thermodynamical processes, as well as the ocean and sea ice to evolve freely. It will be shown that this approach creates accessible, tangible scenarios for everyday weather, heatwaves (on land and at sea), storms, or droughts—that support effective planning and decision-making.
Additionally, it will be shown that spectral nudging in global coupled climate models supports rigorous model evaluation using data from high-quality field campaigns, such as MOSAiC. By aligning model trajectories closely with observed conditions, this approach facilitates direct comparisons between model outputs and field data, strengthening model validation and improvement efforts.

The Korean Peninsula has been experiencing an increasing frequency of extreme weather events in recent years. Despite continuous improvements in Korea Integrated Model (KIM), public satisfaction with weather forecasts remains low. Many experts in Korea attribute this discrepancy to climate change, suggesting that the uncertainty introduced by climate change is outpacing model improvements. In response to these evolving challenges, the KIM is undergoing adaptations to better address climate change impacts. This study aims to maximize KIM's predictive performance in the face of climate change by diagnosing potential performance deterioration and implementing targeted improvements. Two primary areas of focus are: (1) diagnosing and enhancing the effectiveness of climatology used in KIM, and (2) optimizing physical parameterization schemes affected by climate change. This ongoing research is in its initial stages, with preliminary results expected to be available by April at Bonn. The findings will contribute to the development of more robust and adaptive numerical weather prediction models in the context of a changing climate.

This paper presents progress and plans on coupled data assimilation developments conducted at the European Centre for Medium-Range Weather Forecasts (ECMWF) to enhance the exploitation of interface observations in support of climate monitoring and weather prediction.

We describe the ECMWF coupled data assimilation approach. It relies on the atmospheric 4D-Var outer loop cycling methodology which enables to account for parallel minimisations in multiple Earth system components. Increments from each component are applied to initialise subsequent outer loop coupled model trajectories, so assimilated observations have an impact across the Earth system components within each data assimilation window. It is being developed and progressively implemented consistently for land/atmosphere and ocean/sea-ice/atmosphere coupling for operational applications across the ECMWF systems.

Latest developments, results and plans on ocean-atmosphere coupled data assimilation are presented, including work towards coupled sea-ice and sea surface temperature analysis based on passive microwave and infrared radiance assimilation.

We also present progress and plans on coupled land-atmosphere data assimilation conducted in the context of the CERISE (Copernicus Climate Change Service Evolution) and CORSO Horizon Europe Projects. Unified ensemble-based land assimilation combined with outer loop coupled land-atmosphere data assimilation results are shown, along with work to assimilate new observation types over land surfaces, such as Solar Induced Fluorescence, to constrain water and carbon fluxes representations for climate monitoring and forecasting applications.

For both land and ocean coupling, Data Assimilation and Numerical Testing for Copernicus Expansion Missions (DANTEX) activities are introduced and preliminary results shown.

Seasonal wind speed forecasts from the German Climate Forecast System Version 2.1 (GCFS2.1) prediction system, are used to develop the first wind energy forecasts tailored for Germany using a physical model. A subsampling procedure is used on the ensemble forecasts to account for the NAO index and for the three additional modes of sea level pressure variability. The hindcasts for years 1990-2000, with reanalysis data from ERA5 used as a baseline, are used for skill evaluation. Three complementary metrics are used for this purpose: Anomaly Correlation Coefficient (ACC) for deterministic comparison, and Ranked Probability Skill Score (RPSS) and Mismatch for probabilistic comparison. Mid-summer (June-August) and early winter (November-January) seasons in south Germany show a slight tendency for better performance than other seasons in the normal ensemble.The study reveals that the subsampling procedure significantly enhances the skills of the seasonal energy forecasts. This study takes a significant step towards seasonal energy forecasts tools with clear performance indicators that will likely aid in the renewable energy transition.

This talk will provide an overview on current developments in Earth system modelling at ECMWF and summarise challenges and opportunities. It will outline how recent developments in machine learning will change how physics-based models are used in the future, and what future Earth system modelling systems and model development pipelines will need to look like.

Extreme event attribution (EEA) gathers methods from atmospheric sciences and statistics to quantify how exceptional climate or meteorological events can be, and how their outstanding features are connected to climate change. One of the societal expectations of EEA is to provide quick and reliable statements when extreme events occur. Several approaches have been developed towards this endeavor, like the the World Weather Attribution or ClimaMeter, which are the “visible parts of the iceberg”. Although the practical application of fast attribution is debatable for adaptation to climate change, several climate simulation tools have been developed for EEA, which help understand and anticipate the most extreme events, in a changing climate. I will discuss those climate simulation approaches and illustrate them on cases of heatwaves and cold spells in Europe.

Recent climate model developments, established through increased model resolution, have led to improvements in model simulations of the coupled Earth system. However, Earth System models will always produce climate features and variability that differ from the real world and will be prone to biases. Further model improvements are expected to arise specifically from improved representation of physical processes realized through model-data fusion and associated parameter estimation. This will create an unprecedented opportunity to better exploit a large array of Earth observations, from in situ measurements, soil moisture and carbon data, to weather radars and satellite observations, as the resolved scales of the models approach those of the observations. For this, climate data assimilaiton (DA) will be the central tool to bring models and observations into consistency, by improving initial conditions, inferring uncertain model parameters and structure, and quantifying uncertainty. Climate DA must aim to enhance climate knowledge through the improved ability to simulate and predict the real world by optimally combining Earth system models and most available global observations from different Earth system components and domains. In the future, arguably the most important aspects of climate DA and ML in support of model improvement and enhancing predictive capabilities might become optimizing initial conditions, model parameters and model structure to mitigate model biases and thereby improve models’ skill in simulating the observed climate, as well as to enhance model skill for climate projections.

A crucial component of any seasonal forecast system is its retrospective forecasts, or hindcasts, from past years which are used to estimate skill and calibrate predictions. Typically, operational hindcasts cover a span of 20–30 years. Such small hindcast periods are problematic for several reasons. For example, the limited sample size constrains the statistical evaluation of sporadic drivers of seasonal predictability, such as ENSO events, which occur only every few years. In regions with high interannual variability, such as the Euro-Atlantic sector, the limited number of hindcast years is a large contributor to uncertainty for skill estimation and calibration. Furthermore, hindcast skill assessments over extended historical period reveal non-stationarity in skill with fluctuations on multi-decadal timescales. These variations include periods of high skill (e.g., the early 20th century) and low skill (e.g., the mid-20th century), with important implications for seasonal predictability in a changing climate.

In the second part of the talk, I will explore a potential new role for seasonal forecasts in the future. Reliable and successful predictions can provide the testbed for assessing the role of anthropogenic (or otherwise externally driven) climate change for extreme and impactful seasons. As an example, the heavy rainfall during the summer 2022 over Pakistan which led to devastating widespread flooding where thousands of people lost their lives or were injured, will be discussed. Experiments with ECMWF’s seasonal forecasting system SEAS5 with altered CO2 concentrations result in a nonlinear rainfall response to the forcing. A decomposition of the underlying large-scale circulation signals into their atmospheric CO2 and SST-induced responses highlights how the balance of their relative changes controls the future dynamical response. Such forecast-based attribution methods could provide critical information for reliable regional climate adaptation strategies.

Research on atmospheric predictability has historically used physics-based models, which parameterize small-scale processes that strongly influence error growth. Global machine learning (ML) models enable a transformative new approach to predictability research since they have forecast skill comparable to physics-based models at a fraction of the computational cost, and tools to take derivatives of all components of the forecast. We use these tools to map forecast errors from long forecast lead times, when they are large relative to analysis uncertainty, backward in time to the initial condition. This approach yields a deterministic, state-dependent, estimate of the limit of predictability, and statistics on the corrections to ERA5 that may improve long-lead weather forecasts. Results for a large sample show that forecast skill is improved beyond 30 days in all cases. On average, the optimal corrections to ERA5 reflect large-scale adjustments to the tropical atmosphere, especially in regions of deep convection such as the ITCZ.

Land data assimilation aims to monitor the evolution of soil and vegetation variables. These variables are driven by climatic conditions and anthropogenic factors such as agricultural practices. Land surface monitoring includes a number of variables of the soil-vegetation system such as land cover, snow depth, surface albedo, soil water content and leaf area index (LAI). These variables can be monitored by integrating satellite observations into models through data assimilation. Monitoring land variables is particularly important in a changing climate, as unprecedented environmental conditions and trends emerge. Unlike atmospheric variables, land variables are not chaotic per se, but rapid and complex processes affecting the land carbon budget, such as forest management (thinning, deforestation, ...), forest fires and agricultural practices, are not easily predictable with good temporal precision. They cannot be accurately monitored without integrating observations as they become available. Because data assimilation is able to balance information from contrasting sources and account for their uncertainties, it can produce an analysis of terrestrial variables that is the best possible estimate. Data assimilation can involve several techniques, such as 'model parameter tuning', variational assimilation or sequential Kalman filtering methods. The latter are used in meteorology and in some land modelling frameworks to improve initial conditions (e.g. root zone soil moisture) at a given time. New research is being undertaken to assess the impact of improving vegetation initial conditions, as vegetation, like soil moisture, has a memory of past environmental conditions. Vegetation variables such as LAI control the amount of evapotranspiration and their initial conditions have a predictive capability. Examples are given of how data assimilation can be implemented on a global scale by regularly updating model state variables through a sequential assimilation approach. The focus is on LAI assimilation and the use of machine learning techniques to build observation operators that allow the direct assimilation of new vegetation sensitive observations such as microwave backscatter and brightness temperature or solar induced fluorescence. We show that the analysis of LAI together with root zone soil moisture is necessary to monitor the effects of irrigation, drought and heat waves on vegetation, and that LAI can be predicted after proper initialisation. We also show that machine learning can be used to derive new variables (e.g. surface albedo, vegetation moisture) from those already calculated by the land surface model. This paves the way for new developments such as more interactive assimilation of land variables into numerical weather prediction and seasonal forecasting models, as well as atmospheric chemistry models.

Weather forecasting has a long history and has become indispensable for decision-making across various sectors. In recent decades, seasonal forecasts have gained traction due to improvements in predictive skill and the development of sophisticated applications. However, tools addressing the critical time horizon of 10 days to one month – known as sub-seasonal to seasonal (S2S) forecasts – have been less prevalent, primarily due to the limited predictive skill of weather models. The gap is not for a lack of need; there exists significant demand for forecasts beyond the typical weather prediction time frame and shorter than forecasts.

In my work, I have engaged with numerous stakeholders and decision-makers who express a need for S2S forecast information. In this talk, I will present examples from our work at the Climate Futures center in Norway and from past and ongoing climate services projects in Africa.

Motivated by the public and partner requests within Climate Futures, the Norwegian Meteorological Institute has developed a 21-day forecast now integrated into the widely used YR weather app and available through a free API. I will discuss the decision-making process informing the design of this service, and elaborate on how the API is utilized for both commercial and non-commercial products.

In Africa, where reliance on rain-fed agriculture is critical, the timing of the rainy seasons is vital. For farmers, a dry spell following initial rainfall can ruin crops, severely affecting food security and household economics, particularly for subsistence farmers who cannot afford to purchase new seeds. I will detail our ongoing efforts developing "packages" for at-risk communities in Malawi, Ethiopia, and Madagascar. These packages combine 21-day forecasts of dry spells with other resilience-enhancing tools such as training, insurance, cash support, and soil moisture instruments, aiming to improve community robustness to weather fluctuations.

In summary, we see vast potential for S2S products and services to significantly benefit both Europe and Africa, addressing critical needs and fostering resilience in diverse communities.

The stratosphere represents an important source of predictability for the extratropics in winter and spring, especially after extreme stratospheric events. These events tend to be followed by persistent temperature and precipitation anomalies at the surface, including extreme weather events. While the influence of the stratosphere on the troposphere has long been established, the predictability arising from this downward influence is less straightforward to quantify due to the strong case-by-case and geographic variability that models cannot fully reproduce. It further remains unclear how the coupling between the stratosphere and the surface may have changed over time in the recent past and how it will change under climate change. Such changes may affect both the dynamical vertical coupling as well as the predictability arising from stratospheric sources. This contribution will provide an overview of dynamical stratosphere - troposphere coupling and its impacts on extended-range prediction.

Sub-seasonal prediction has been operational at ECMWF since 2004. Over the past 20 years, the forecast skill at week 3 and 4 has significantly improved, particularly in the Tropics, thanks to improved modelling, model complexity, initialization and model ensemble configuration. However, further progress is impeded by the presence of model biases which affect the prediction of key sources of sub-seasonal predictability, their interactions and their impacts. For instance, despite significant improvements in the representation and prediction of the Madden Julian Oscillation since 2004, the representation of its impacts in the tropics and Extratropics in S2S models remains problematic and a major obstacle for further skill improvement at weeks 3 and 4 over Europe. Machine learning represent an opportunity to overcome these issues by providing improved post-processing, replacing the dynamical model, helping correcting models errors or helping better understand causality at the S2S timescale.

TBC

National Center for Environmental Prediction (NCEP)’s Environmental Modeling Center (EMC) is a premier developer of Numerical Weather Prediction (NWP) systems and is responsible for transitioning and maintaining more than 20 numerical prediction systems for use across the NCEP and the rest of the National Weather Service (NWS) and NOAA. These systems are developed through close collaboration with an extensive network of partners from the academic, federal and commercial sectors. Together with the other centers in NCEP, EMC is a critical resource in national and global weather prediction and is a starting point for nearly all weather, water and seasonal climate forecasts in the United States.

NOAA’s Next Generation Global Prediction System (NGGPS) Project has initiated a major shift in the development of operational NWP systems with a goal to simplify the NCEP Production Suite (NPS) using the Unified Forecast System (UFS) framework. EMC has taken a lead in developing and consolidating various operational systems into UFS based applications. The UFS is designed as a community-based, comprehensive atmosphere-ocean-sea-ice-wave-aerosol-land coupled Earth modeling system with coupled data assimilation and ensemble capabilities, organized around applications spanning local to global domains and predictive time scales from sub-hourly analyses to seasonal. Various operational applications that are currently developed and maintained by EMC in support of various stakeholder requirements are expected to transition to the UFS framework in the next few years, and share a set of common scientific components and technical infrastructure for consolidating NPS applications and to accelerate transition of research into operations.

This talk describes major development and operational implementation projects at EMC, how those fit within the broader strategic plans of NOAA, and how these projects link with other model-related projects internally within NOAA and with the broader U.S. and international modeling community.

Data assimilation is the process that connects observations to numerical models to produce initial conditions for prediction. Advancements in data assimilation have been some of the foundations for realizing improved predictions from operational numerical weather prediction over the past several decades. This presentation will provide an overview of some of the major advancements in data assimilation that have led to drastically improved operational predictions. These include developments to perform direct assimilation of satellite radiances (instead of meteorological retrievals) and the adaptive bias correction thereof, extensions to 4-dimensional assimilation including techniques such as 4DVar, and incorporation of ensembles to produce flow-dependent estimates of uncertainties. An overview of the current state of the science for data assimilation at operational numerical weather prediction centers will be presented. Finally, the possibilities for the future will be discussed, including the prospects for moving toward systems based entirely on artificial intelligence.

As the national meteorological and climatological service MeteoSwiss has the mandate to provide weather and climate services for the safety and benefit of Switzerland. Being one of the founding members in 1975, Switzerland uses global medium-range weather forecasts and climate reanalysis since the beginnings of ECMWF. However, the country’s complex orography and local weather phenomena in the Alps impose a special challenge in providing accurate and timely local weather and climate information across scales.
In this lecture I will present examples of seamless weather and climate services tackling the challenges imposed by complex orography. An emphasis will be laid on products using ECMWF data as input. Some of the latest developments concern seamless workflows for weather and climate monitoring, wind climatology and the provision of a national drought platform.

Quantifying changing risks of extreme weather has traditionally been seen as a climate problem: begin with a scenario, run a global climate model, downscale it. This inevitably creates a tension between the reliability of the model, to capture the event-types of interest, and ensemble size, to capture the event-probabilities of interest. We explore a complementary approach, perturbing the initial and boundary conditions of an ensemble forecast to assess the impact of external drivers on extreme weather risk. The advantage of this approach is that it starts from a probabilistic simulation of the event in question whose reliability is already known, so we avoid making climate statements about weather events that the model is unable to simulate. It also takes advantage of the fact that the impact of external drivers on extreme weather, even on decadal timescales, will often be significantly less than the spread of a medium-range forecast, so it makes more sense to start from a reliable ensemble and apply an uncertain (but small) perturbation than to start from an unreliable model and apply a more realistic perturbation. We demonstrate this approach on some examples of recent extreme weather events, and show that results are broadly consistent, albeit often more conservative, than other approaches to the extreme weather trend-attribution problem. We argue that ensemble forecasting systems, including medium-range reforecasts, are an underappreciated tool for climate research, particularly relevant to near-term adaptation decisions.

Data-driven models (DDMs) are rapidly emerging as an alternative to Numerical Weather Prediction (NWP) models. The success of global DDMs is driving increased efforts towards tailoring DDMs also for regional applications. National meteorological and hydrological services, such as MET Norway, primarily depend on high-resolution regional NWP models for delivering weather forecasts to the public and industries affected by weather conditions. Given Norway's complex topography and coastline, forecasts must be provided at a higher spatial resolution than what is commonly provided by global DDMs.

A DDM suitable for regional weather forecasting applications is presented. This model extends the Artificial Intelligence Forecasting System (AIFS) by integrating a stretched-grid architecture. This architecture allows for a higher resolution in targeted regional areas, while maintaining a lower resolution globally. The model is based on graph neural networks, which support various multi-resolution grid configurations. Using this framework we will discuss the potential of machine learning in addressing the multiple challenges associated with delivering location-specific forecasts to end users. These include generating forecasts on kilometer scale with hourly temporal resolution, incorporating other observation sources such as extensive networks of crowd-sourced observations, providing probabilistic forecasts, and forecasting rare events.

The Madden-Julian oscillation is one of the most important sources of subseasonal predictability in the mid-latitudes as it modulates the occurrence of long-lived and quasi-stationary weather regimes through teleconnections. This modulation of the mid-latitude flow in turn has consequences for the activity of low pressure systems in the mid-latitudes, which themselves can significantly influence the dynamics and predictability. By evaluating reanalysis data as well as idealized experiments, we show in the first part of this study that the release of latent heat in midlatitude warm conveyor belts (WCBs) significantly modifies known canonical teleconnections or even suppresses them. In state-of-the-art subseasonal forecast systems, the modulation of the frequency of WCBs conditioned on the MJO state is erroneously represented. Accordingly, we hypothesize that an improved representation of this link could have a positive impact on subseasonal weather regime forecasts during active periods of the MJO. In the second part of the study, we investigate the potential of artificial intelligence-based weather prediction models to identify source regions of forecast errors that are particularly detrimental to the subseasonal forecast skill. Using the hybrid Neural GCM, we find that the model is similarly sensitive to different relaxation regions as reference experiments with numerical weather prediction models.

Machine learning (ML) is transforming the field of weather and climate forecasting, offering innovative approaches to enhance accuracy, efficiency, and utility. This presentation provides an overview of how ML is being integrated into forecasting workflows, highlighting its transformative potential and diverse applications.

We will examine ML's role in improving forecast skill, from emulating complex physical processes to accelerating data assimilation and building ensemble systems. ML facilitates the exploitation of vast Earth observation datasets, perhaps enabling the use of even more Earth-system data in developing and running forecasting systems.

In the context of the 50th anniversary, we will look out into the future, and try to forecast the role for machine learning in weather and climate in the years to come.

AI-driven weather forecast models are now more accurate and much faster than the best physics-based models. Can similar technology be used to develop AI climate models for reliably simulating decades to centuries across a range of climates? Groups around the world are making exciting progress toward this goal using a range of approaches. This talk focuses on the open-source Ai2 Climate Emulator (ACE). ACE can be trained to accurately emulate both daily weather variability (including extremes) and climate of historical reanalysis (ERA5), or of a global atmospheric model,. ACE has 100-1000x smaller computational cost of running a physics-based model of comparable grid resolution. ACE can also accurately emulate the weather and climate of finer-grid physics-based models. ACE can be forced by specified sea-surface temperatures (AMIP) or coupled to a slab ocean model; coupling to an AI emulator of a dynamical ocean model is ongoing. When trained on reference model simulations spanning multiple climates forced by changed CO2 concentrations, ACE can accurately emulate the climate change response of the reference model. Lastly, we note some key remaining challenges to general-purpose use of ACE for climate modeling applications.

AI applications to weather and climate are growing exponentially and there is a critical need to ensure that these methods are trustworthy before they are deployed. In this talk, I will describe work from the NSF AI Institute for Research on Trustworthy AI in Weather, Climate, and Coastal Oceanography (AI2ES) focusing on studies with forecasters to determine what they need for an AI forecast to be trustworthy. We discuss the elements of trust as seen by the forecasters and how this affects the AI development lifecycle. Finally, I end with a discussion of developing trustworthy AI model verification via Extreme Weather Bench.

Eastern Africa experiences frequent extreme weather/climate events, such as droughts, floods, and heatwaves, which are projected to intensify due to global warming. To address these challenges, a variety of forecasting techniques have been developed to predict weather and climate conditions across different timescales. These methods include dynamical, statistical, and hybrid approaches, each tailored to meet the unique needs of the region and users.
For weekly timescales, dynamical downscaling techniques are employed to produce tailored forecasts, such as exceptional rainfall events, temperature and rainfall anomalies, and mean conditions. At sub-seasonal to seasonal timescales, statistical downscaling methods are used to downscale outputs from global climate models, providing actionable information for stakeholders in critical sectors like agriculture, health, and disaster risk reduction. Since 2019, the IGAD Climate Prediction and Applications Centre has been utilizing objective forecasts, which are useful to drive the impact models and produce tailored forecasts.
In recent years, machine learning has emerged as a promising tool in the region's forecasting efforts. Conditional Generative Adversarial Networks (cGAN) have been developed for short-range timescales, enabling National Meteorological and Hydrological Services (NMHSs) to objectively provided probabilistic forecasts and calculate probabilities of exceeding specific weather thresholds. In addition to the short-range forecasts, machine learning techniques have also been developed for the seasonal timescales. These advancements are likely to significantly improve the early warning systems over the region.
By integrating diverse forecasting techniques and advancing machine learning capabilities, Eastern Africa is building a more robust and adaptive forecasting system, essential for mitigating the impacts of extreme weather events and enhancing climate resilience in the face of a warming planet.

El Nino-Southern Oscillation is a major mode of climate variability at interannual time scales, with worldwide impacts in weather patterns. This lecture will provide an overview of the processes influencing ENSO, their positive and negative feedbacks and the interaction between different time scales. It will discuss the implications for ENSO forecasting and the impact of climate change. The discussions will be illustrated by the comparative analysis of major recent El Nino events.

The Korea Institute for Atmospheric Prediction Systems (KIAPS) was established in 2011 with the objective of developing a global atmosphere-only numerical weather prediction system for operational use at the Korea Meteorological Administration. The system was completed in accordance with the established timeline and became operational at KMA in April 2020, demonstrating immediate and exceptional performance. The system is based on a novel atmospheric model called the "Korean Integrated Model" (KIM), which is structured on a cubed-sphere grid and employs the spectral element method within its dynamical core. The deterministic data assimilation (DA) process is based on a hybrid-4DEnVar algorithm, while the forecast uncertainties are modelled by a 50-member Ensemble Prediction System (EPS). The EPS is based on a local ensemble transform Kalman filter (LETKF) data assimilation (DA) algorithm, and further schemes are included to account for deficiencies and uncertainties in the DA process and the forecast model. Ensemble forecasts are contingent upon the manner in which initial uncertainties and model uncertainties are represented. This study introduces the stochastic representation schemes of model uncertainties in KIM and presents the results of numerical experiments. We will provide a preliminary examination of the results, with a particular focus on the sensitivities of the model to stochastically perturbed physical tendencies (SPPT) and stochastically perturbed parameterizations (SPP) schemes. To reduce computational costs, sensitivity tests were performed in a low-resolution framework to analyze the impact of these schemes on the ensemble mean error and spread in temperature, geopotential height, and specific humidity in extended medium-range forecasts. This will followed by a discussion of possible future plans to further enhance the methods used to perturb the forecast model.

Advances in pre-exascale computing and model scalability have enabled the use of Earth system models with grid spacing of a few kilometres (km-scale models) to forecast weather and simulate climate over decades. Km-scale models provide more accurate local data and better capture the physical processes conducive of extreme weather events and interactions between model components, which coarser models often miss.

This presentation review the projects at ECMWF that have made possible the development and implementation of km-scale models and assess their current performance. Key achievements include better representation of extreme weather events, like tropical cyclones and orographic precipitation, and improved predictability of large-scale atmospheric patterns.

We also address remaining challenges, such as improving the parameterization of moist physics at the km scale for greater physical realism without reducing NWP skill, advancing non-hydrostatic models, enhancing data assimilation for km-scale resolution, and quantifying uncertainty despite high simulation cost.

To fully exploit km-scale information, application models must be seamlessly integrated into their workflow. This integration is a fundamental aspect of the Digital Twins for weather and climate developed in Destination Earth, enabling the delivery of more relevant, impact-oriented and actionable information.

The production of ECMWF's latest atmospheric reanalysis, ERA6, will begin in 2025. ERA6 benefits from 8 years of research and development since the start of it's predecessor, ERA5, encompassing improvements to the forecast model, data assimilation system and the exploitation of observations. The horizontal resolution, at 14km (TCo799), represents a factor of two improvement over ERA5 (30km, TL639). Of particular relevance for ERA6 is the improved treatment of biases in the stratosphere through the use of weak constraint 4DVar and the introduction of outer-loop coupling of ocean and atmosphere. The provision of ocean forcing from ECMWF's latest ocean reanalysis, ORAS6, represents a significant improvement over ERA5. ERA6 will also benefit from the assimilation of many reprocessed datasets, both conventional and satellite, which have been the subject of extensive testing in recent years. Based on pre-production testing the step change in re-forecast quality of ERA6 relative to ERA5 is similar to that achieved in the the upgrade from ERA-Interim to ERA5. It is expected that the initial production streams, covering the period from 2006-present, will be published in the second half of 2026.

The rapid evolution of societal needs, driven by climate change, technological advancements, and shifting user expectations, presents significant challenges for traditional weather services. In response to these developments, weather services need to transform from traditional forecast providers to becoming decision-making consultants that empower diverse and also new user groups.
The presentation examines the transformation required to address this paradigm shift, focusing on the integration of sector-specific expertise and developments in co-design with key customers. This includes discussing critical hurdles such as rapidly evolving customer expectations versus the existing pace of internal renewal. This will lead us to examine organizational and personal challenges, highlighting strategies for fostering innovation, collaboration, and continuous renewal in weather services.

Forecasts on sub-seasonal to seasonal (S2S) timescales have huge potential to improve early warning and anticipatory action ahead of high impact events. However, fully realising this potential predictability requires reliable forecasts that are communicated effectively so that they can support appropriate preparedness action. I will reflect on the African SWIFT (Science for Weather Information and Forecasting Techniques) S2S forecasting testbed which brought together researchers, forecast producers and forecast users from a range of African and UK institutions. The testbed used a co-production approach to pilot the provision of real-time bespoke S2S forecast products for applications. The S2S testbed supported decision-makers in a range of sectors and contexts. For example, informing food security decisions and hydropower energy planning in East Africa, supporting agricultural decision-making across West Africa, and, in health applications, increasing the lead-time for potential disease outbreaks.

I will critically reflect on the benefits and challenges of the co-production process within the S2S applications context. Specifically, while having direct access to the real-time S2S data allowed user-guided iterations to products to make them more actionable for their specific context. Some key lessons for effective co-production emerged. First, it is critical to ensure there is sufficient resource to support co-production, especially in the early co-exploration of needs. Second, all the groups in the co-production process require capacity building to effectively work in new knowledge systems. Third, evaluation should be ongoing and combine meteorological verification with decision-makers feedback. Ensuring the sustainability of project-initiated services within the testbed hinges on integrating the knowledge-exchanges between individuals in the co-production process into shaping sustainable pathways for improved operational S2S forecasting within African institutions.
I will also talk about opportunities to develop this thinking in two new projects: 1) ACACIA (Anticipatory Climate Adaptation for Communities in Africa) and 2) POWER-Kenya (Potential of sub-seasonal for Operational Weather and climate information for building Energy Resilience in Kenya).

Climate models show a warming trend in the eastern Tropical Pacific that is inconsistent with observations, an error which has consequences for the estimation of climate sensitivity. Similar trends are detected in hindcasts of ECMWF’s operational seasonal forecasting system SEAS5 from 1993 onward. Results show that SEAS5 hindcasts strongly underestimate the observed strengthening of equatorial easterly and cross-equatorial southerly winds. Associated with this, hindcasts exhibit spurious trends towards anomalously warm sea surface temperature (SSTs) in the equatorial and southeastern tropical Pacific, consistent with a tendency towards overprediction of probabilities for El Niño towards the end of the hindcast period. This tendency persisted during the 2020-2022 triple-dip La Niña, with forecast errors reminiscent of the long-term trend errors. Uncoupled hindcasts with prescribed SSTs exhibit similar wind trend errors as SEAS5, suggestive of an underestimation of the wind-evaporation-SST feedback by the atmospheric model. Spurious cloud and shortwave radiation trends contribute to SST trend errors as well. Results also indicate that a persistent positive precipitation bias in the Western Pacific warm pool keeps the model in a regime where equatorial Pacific winds are too insensitive to changes in zonal gradients. In coupled mode, trend errors of the atmospheric model amplify through the Bjerknes feedback. The excessive warming of the equatorial ocean subsurface in the ocean initial conditions also plays a role. Further reduction of model biases are thus needed to tap the source of predictive skill represented by climate trends, and seasonal forecasts are an efficient test bed for climate model development.

Sooner or later all forecasts will be wrong. This is given by the nature of the atmosphere being chaotic. A chaotic system is sensitive to initial condition errors meaning that if the prediction starts to differ from the state of the nature, the evolution the two will eventually be uncorrelated even if the initial errors are very small. How these differences growth is highly situation dependent. The error growth is also affecting the smaller scales much faster than longer scales.

However, there are also other contributors to the forecast errors. This includes approximation in the model or missing processes, spatial representation differences between the model and the anticipated forecast, and interpretation errors. Finally in long-range forecasts, model errors are playing an important role in creating errors. In this presentation we will discuss the different aspects of the forecast errors in the light of severe weather predictions, and also give some perspective of the future in terms of predictability related to data-driven forecasting.

After more than quarter of century of operational physical-model-based seasonal forecasting, seasonal prediction remains a challenging problem. The basics of the forecast problem are well understood, and at least in recent decades we can initialise our forecasts moderately well. Many aspects of variations in seasonal weather can be predicted with a demonstrable level of skill. And yet. Although there have been steady improvements in model fidelity, the pace of progress on critical problems has been slow, relative to what is needed for forecasts to be credible and reliable. As a consequence, predictability limits, which are most conveniently studied by model experimentation, are poorly known.

How might progress in seasonal forecasting be accelerated? Can we create realistic physics-based models of the relevant parts of the earth’s climate system, given computer constraints? Can we improve reanalyses of past decades? And can we pool information from the past and present, knowledge of climate forcings, physical models and machine-learning techniques to transform what seasonal forecasting is capable of?

ECMWF has developed a machine-learned weather prediction model, the Artificial Intelligence Forecasting System (AIFS). We will describe the current state of AIFS, its architecture and framework, and discuss ensemble generation methods, via probabilistic score optimisation and diffusion-based training. AIFS produces highly skilled forecasts ranging from medium-range to sub-seasonal time scales. Both, a deterministic AIFS version as well as an ensemble version of AIFS are now run in experimental mode alongside ECMWF’s physics-based NWP model. AIFS forecasts are available to the public under ECMWF’s open data policy.

Climate services can be considered as the provision and use of climate information in decision-making. As with weather services based on weather information, climate information has long been used to inform decisions. However, the development, delivery and use of climate information became a global endeavour in the wake of World Climate Conference-3 in 2009 when heads of state, government ministers, industry representatives, and scientific and technical experts from many fields of practice recognised the growing importance of climate services. We have experienced huge developments across the climate service value chain, including the underpinning observations, research and modelling capability providing the climate information, through to the service delivery and uptake for decision-making. But of course, more needs to be done, particularly as we continue to see climate-related events impacting society and economies worldwide against the backdrop of climate change. This presentation will introduce climate services, considering some history, the landscape, the drivers, and developments, highlighted with examples. The presentation will also highlight current and future challenges.

Subseasonal to decadal prediction science and services have advanced rapidly in recent years, with the WCRP-involved S2S project and Grand Challenge in Near-term Climate Prediction having laid key groundwork for the establishment of WMO services via its Lead Centres for Sub-seasonal Predictions Multi-Model Ensemble (SSPMME) and Annual-to-Decadal Climate Prediction (ADCP), respectively. Ongoing WCRP groups and initiatives such as its Working Group on Subseasonal to Interdecadal Prediction (WGSIP), Decadal Climate Prediction Project (DCPP) and Explaining and Predicting Earth System Change (EPESC) are further exploring scientific frontiers across the spectrum of prediction time scales between those of numerical weather prediction and long-term projections. This presentation will briefly summarize these activities within WCRP, and then share multifaceted perspectives on two broad frontier topics: predicting a rapidly changing Earth system, and the AI/ML revolution—where will it lead?

The DestinE Digital Twin for Climate Change Adaptation (Climate DT) supports adaptation activities through the provision of innovative climate information on multi-decadal timescales, globally, at scales at which the impacts of climate change are observed. This initiative presents the first ever attempt to operationalise the production of global multi-decadal climate projections at resolutions of 5 to 10km, leveraging the world-leading supercomputing facilities of the EuroHPC Joint Undertaking along with some of the leading European climate models.

This lecture will give an overview of the Climate DT system, including the underlying models, the novel technological approaches, and data handling and analysis. It will also showcase how the model output is transformed into climate information by using impact sector models as part of the Climate DT.

https://destine.ecmwf.int/climate-change-adaptation-digital-twin-climate-dt/

Operational ensemble forecasting with NWP models started over 30 years ago. Nowadays, ensembles are in use at all forecast lead times and they are the established source for probabilistic forecast information. The talk begins with a look at the evolution of the skill of the ECMWF medium-range ensemble and how that compares with the skill of the unperturbed, deterministic forecast.

Then, an overview follows of the representation of model uncertainties used in the IFS, past and present. This includes a recent major revision: In 2024, the Stochastically Perturbed Parametrization scheme (SPP) was implemented. It is used in medium-range forecasts, sub-seasonal forecasts, the next version of the seasonal forecasts (SEAS6) and the ensemble of data assimilations (EDA) including the EDA of the upcoming reanalysis ERA6. The advantages of representing uncertainties with SPP and areas for further improvement will be summarised.

Lastly, an overview is given on the representation of initial uncertainties, past and present as well as ongoing work that explores the introduction of a scale-dependent re-centring of the EDA on the deterministic analysis.

Land surface albedo is crucial for modeling energy exchange between the land and atmosphere. It determines how much sunlight is reflected under cloud-free conditions, influencing surface temperature, evaporation, snow and ice melt, and heat exchange processes.

A common approach to estimating land surface albedo involves assigning intrinsic values based on wavelength, direction, and categorized surface types. However, surfaces often consist of complex mixtures of components, and geometric topography adds further complexity through shading effects. MODIS-based albedo accounts for these complexities directly in its values.

In 2015, we introduced in KIM (Korean Integrated Model) a MODIS-based albedo parameterization using a climatological approach, leveraging 16 years of MODIS satellite data from 2000 to 2015. Since then, Earth's surface has undergone noticeable changes due to factors like land use, wildfires, and vegetation shifts.

To address these changes, we updated the dataset using MODIS data from 2021 to 2023, supported by enhanced data processing algorithms.This update highlights the importance of accurate climatological data in understanding and assessing the impacts of climate change.

Ocean and sea-ice ReAnalyses (ORAs) are reconstructions of historical ocean and sea-ice states generated by incorporating observations into simulated model states through data assimilation methods. ECMWF has a long history of development and production of ORAs. Initially conceived to provide initial conditions for seasonal re-forecasts, the ECMWF ORAs now support climate monitoring and the initialisation of all ECMWF coupled forecasts (medium-range, sub-seasonal and seasonal forecasts). ORAS6 - the 6th generation of ocean and sea-ice ensemble reanalysis developed at ECMWF – has an important addition: the provision of ocean and sea-ice conditions for the upcoming reanalysis ERA6. ORAS6 features numerous upgrades compared to ECMWF’s current system-5 (ORAS5). The ocean and sea-ice model has been upgraded and is now driven by hourly ERA5 forcings. A novel ensemble-based variational data assimilation system has been developed. This new EDA system is constructed with a hybrid B covariance model that provides flow-dependent background error variances and correlation scales. Other updates include direct assimilation of L4 SST and L3 sea-ice concentration data, a revised method for altimeter data assimilation, and updates in the bias correction scheme. The performance of ORAS6 is significantly improved compared to its predecessor ORAS5, especially for the estimation of the upper ocean and ocean transports. This presentation aims to provide an overview of ocean and sea-ice reanalysis activities based on the ECMWF system. The impact on coupled ocean-sea-ice-atmosphere forecasts initialized from ORAS6 will be discussed as well.

The Meiyu rainfall is crucial for Taiwan's water supply, often marking the first significant influx of water following a prolonged dry season. However, Meiyu also brings extreme rainfall. For example, from June 2 to 4, 2017, torrential rains embedded within a quasi-stationary Meiyu front brought extreme rainfall to Taiwan. Understanding the mechanisms and variations of the Meiyu is critical to weather forecasting and climate projection in East Asian countries. Ding et al. (2020) reviewed the multi-timescale variabilities of Meiyu related to impacts from both quasi-biweekly and 30–60-day oscillation.
Using wavelet analysis, we selected 2017, characterized by short-period oscillations, and 2007, characterized by long-period oscillations, as case study subjects to explore the impact of different periods on Taiwan's Meiyu season. We used ERA5 as the initial and boundary conditions for the model, employing a Butterworth filter to extract short-period (ISO-S) and long-period (ISO-L) components, and ultimately used the Cloud-Resolving Storm Simulator (CReSS) to reproduce and simulate Meiyu cases under different periodicities.
The results show that ISO-S causes the front to move more slowly and become structurally looser, transforming the originally concentrated and intense convective cells into weaker and more dispersed ones, thereby leading to rainfall occurring across the entire Taiwan during the Meiyu front period. The strengthening of the southwest monsoon flow might be one of the primary reasons. By analyzing the water vapor budget equation, we found that ISO-S generally increases rainfall across different regions of Taiwan. The primary mechanism is influenced by the Convergent Vapor Flux (CVF), which is predominantly driven by convection (CONV), though environmental dry advection (ADV) partially counteracts this increasing trend.
In the future, we will also analyze the case of the 2007 Meiyu front, which was dominated by long-period oscillations, to discuss the impact of strong ISO-L on the Meiyu in Taiwan.

The potential use of European Centre for Medium-Range Weather Forecast (ECMWF) ensemble prediction system SEAS5 over Mainland Southeast Asia was evaluated. The evaluation spans 30 years (1985–2014), examining SEAS5's skill in combination with the Variable Infiltration Capacity (VIC) hydrological model for streamflow forecasts, as well as the WOrld FOod Studies (WOFOST) crop model for rice production forecasts. These hydrological and agricultural results were compared against the WFDE5-driven reanalysis using verification skill metrics at grid cells for each month. The findings reveal predictive capabilities for streamflow forecast extend to a 1-month. Noteworthy, strong seasonal and regional dependence occurs, with high forecast skills during the pre-monsoon (April–May) and post-monsoon (October–November). The significant streamflow skill at each initiation month and lead time corresponds to the forecasting skill of meteorological variables. For the rice prediction, SEAS5 exhibits high performance at the beginning of the rainy season, where strong seasonal climate predictions are observed. The model shows the ability to capture anomalous rice yields and consistent accuracy throughout a 1-month to 3-month forecast. However, limitations in skill are evident when rice planting times are delayed by one or two months during the rainy season, as well as when planting in the dry season. SEAS5 shows useful skills that can potentially be used for hydrological and agricultural anticipatory management. The results could already support an initial step to come to potential anticipatory (agro-)hydrological management and could be utilised as an input for an early warning system in various sectors.

Climate change has been linked to tropical cyclone (TC) poleward migration, extreme precipitation, and an increased proportion of major hurricanes. Understanding past TC trends and variability is essential for projecting future impacts on human society. However, limited observational data, subjective analyses, and the spatiotemporal inconsistencies in traditional best-track datasets have hindered confidence in assessing TC responses to a changing climate. To address these challenges, this study employs deep learning to reconstruct past "observations", creating an objective, homogenized global TC wind profile dataset for 1981–2020.

The proposed model converts multichannel satellite imagery into 0–750-km wind profiles of axisymmetric surface winds, enabling a consistent representation of TC intensity (Vmax), structure, and integrated kinetic energy (IKE). Trained on labeled data combining best tracks and numerical model reanalysis of 2004–2018 TCs, the model demonstrates strong performance verified against independent satellite-derived radar surface winds, making it suitable for climate analyses.

The new homogenized dataset reveals significant trends: a ~13% increase in the proportion of major TCs over the past four decades and a ~25% increase in extreme-IKE TCs (IKE > 135 TJ, the 95th percentile). Additionally, the mean IKE of high-IKE TCs (IKE > 79 TJ, the 80th percentile) shows an increasing trend exceeding one standard deviation of 40-year variability. By introducing a novel approach to reconstructing historical TC "observations," this deep learning-based dataset offers unprecedented insights into TC structural extremes and provides a valuable tool for validating climate simulations.

The Korean Integrated Model (KIM) was developed for global numerical weather prediction and has been operational in real-time since April 2020. To expand its capabilities from deterministic medium-range weather forecast to probabilistic sub-seasonal to seasonal (S2S) prediction, it is essential to improve the representation of physical processes and the interactions between the atmosphere and surface. To address this, the Korea Institute of Atmospheric Prediction Systems (KIAPS) is developing a new coupled modeling system by advancing the land surface processes and coupling the ocean, sea ice, wave, and river-routing models, along with the further development in ensemble forecast and coupled data assimilation.
In the new coupled modeling system, the Noah land surface model (LSM) has been replaced with the Noah with multi-parameterization options (Noah-MP) LSM version 5, enhancing the biophysical and hydrological complexity of the land surface processes. The Nucleus for European Modelling of the Ocean (NEMO) version 4, including the Sea Ice Modelling Integrated Initiative (SI3), has been coupled to represent atmosphere-ocean-sea ice interactions. The system has also integrated the WAVEWATCH III (WW3) version 7 to account for ocean wave effects and the Catchment-based Macro-scale Floodplain (CaMa-Flood) version 4 to manage land-to-ocean runoff. Additional efforts have been made to ensure physical consistency among the components, with improvements in the parameterizations for the surface layer and skin temperature.
The latest version of the fully coupled model demonstrates comparable performance on medium-range weather forecasting and seasonal simulations compared to the atmosphere-only model. In the workshop, the current progress and future plans will be presented in detail.

Tropical Cyclones (TCs) are a frequent extreme weather phenomenon that often causes severe damage in the Northwest Pacific region, making accurate forecasts important. A typical approach is using Numerical Weather Prediction (NWP) because its forecast performance has continuously improved over the past few decades. However, the high computational cost remains a major obstacle to progress. Therefore, we proposed using PanGu Weather (PGW), a machine learning model trained on reanalysis data, to reduce computational costs while retaining forecast accuracy for global metrics and extreme events. However, there are still some underlying problems between the basic dynamic balance relationships in the atmosphere, and its forecasting ability cannot surpass the current NWP ensemble forecasts. To address these issues, we perform ensemble forecasts of PGW, and use common forecast verification tools to evaluate the performance of single-member forecast and ensemble mean forecast of PGW in predicting TCs' track, intensity, and other meteorological variables compared to observation data.
In each TC case, the ensemble mean forecast of PGW shows higher forecast accuracy in track than the single-member forecast of PGW in both the long and short lead time. At the same time, the same trend can only be seen in the short term (1-3 days) for variables like geopotential height, winds, and mean sea level pressure. Interestingly, we also identified that both ensemble mean forecasts and single-member forecasts of PGW have high accuracy in shorter lead times in terms of TC intensity forecasts. We hypothesize the low accuracy shown in the long term of both of the forecast types is due to the inherent physical fitting errors in the PGW model.

Recently, weather prediction models based on artificial intelligence (AI) have become equally to slightly more accurate than the established operational weather prediction models in terms of deterministic scores. The much lower computational cost of the AI-models facilitates the generation of large ensembles, hence it is important to assess if their error growth properties are realistic. Here, we investigate uncertainty growth from initial condition perturbations of varying amplitudes, simulated with AI-weather prediction models (PANGU, GraphCast, FourCastNet, Neural-GCM, Aurora, Gencast) and with a “classic” fluid equation-based weather model (ICON). From past research and the global convection-permitting ICON simulations, it is expected that small-amplitude initial condition perturbations would grow very fast initially in areas with latent heat release, then spreading out to larger and larger scales, eventually setting a fixed and fundamental limit to the predictability of weather. This phenomenon is known as the butterfly effect. We find however, that in contrast to ICON, the AI-based models completely fail to reproduce the rapid initial growth rates. Instead their growth rates remain similar to those typically found on synoptic-scales, which incorrectly suggests an unlimited predictability of the atmosphere. In contrast, if the initial perturbations are large in amplitude and comparable to current uncertainties in the estimation of the initial state, the AI-based models basically agree with results from the “physically-based” ICON simulations, although some deficits are still present, mostly related to their particularly low effective resolution. This work provides an example of how machine learning models can fail to reproduce a fundamental physical principle, even though they can accurately mimic many observed behaviors.

Within the context of the “Global-to-Regional ICON digital twin” (GLORI) project, a convection-permitting ensemble forecasting is established in order to study the predictability of high-impact weather events with high-resolution modeling (up to 500 m) and the influence of the land-surface—atmosphere coupling mechanisms.
At the DWD, ICON-D2-EPS is the limited-area high-resolution component of the ICON modeling system, running as an ensemble of 20 members at 2 km horizontal resolution over Germany and surrounding areas. The perturbed initial conditions are provided by the km-scale ensemble data assimilation system KENDA, run at the same resolution, assimilating a wide range of observations, including radar-derived radar volumes. Boundary conditions are provided by ICON-EPS, the global ensemble with a refinement at 13 km over Europe and is refreshed every 3 hours.
In this work, we employ a nested domain with horizontal resolution of 1-km in the southern region of the ICON-D2, encompassing the Alps mountains. We run a 24-hour forecast simulation starting at 00UTC on the 21st of June 2022, with 20 ensemble members. The choice of the date is crucial as it corresponds to a day when the DWD recorded instances of heavy rain and hail in southern Germany.
In our study, we perform all experiments using a two-moment microphysics scheme. Additionally, we incorporate the standard operational model perturbations and subsequently analyze the influence of various convection schemes on the predictability of processes that lead to convection development. Specifically, we examine the behavior of the convection scheme in two configurations: shallow convection only and deep convection parameterization in the so-called gray-zone-tuning version. By selectively enabling and disabling these schemes, our goal is to evaluate their individual contributions to predictability.
Following this, we implement a tailored variant of the stochastically perturbed parameterization scheme (SPP) in ICON in order to delve into the influence of some uncertain parameters within either microphysics or turbulent parameterization, further advancing our understanding of its effects on the model performance.

A budget analysis framework is proposed to quantify contributions of various uncertainty sources to total variance within an ensemble forecast. The method evaluates impact of uncertainty sources by successively incorporating their representations and taking derivatives to measure their impact. The framework decomposes total ensemble variance into sum of individual variances attributed to each uncertainty representation and their interaction terms in a form of correlation.

The decomposition technique is applied to ensemble variance of limited-area forecasts, including operational initial and boundary condition (IBC) uncertainties. These are further augmented by two model uncertainty representations: the physically-based stochastic perturbations scheme (PSP) and microphysical parameter perturbations (MPP).

When analysing grid-point variance, PSP and MPP variances show similar growth patterns, which are strongly linked with the diurnal cycle of local convective activity. Additionally, their interactions are characterised by systematic negative correlation with other impact, suggesting a large degree of double counting. Extending the analysis to scale-dependent variance reveals that that the double counting comes from feature displacements at small scales, and, when accounting for interactions, PSP and MPP introduce net variability on different scales. These results encourage consideration of scale-dependent interactions between uncertainties for accurate evaluation of new uncertainty schemes.

The impact of two stochastic parameterization schemes is compared in a convective scale model. The implementation
of the physically based stochastic boundary layer perturbation (PSP) scheme in the Weather Research and Forecasting
(WRF) model represents uncertainties arising from unresolved boundary-layer processes due to the finite grid size.
Uncertainty in the microphysical processes is represented by the stochastic parameter perturbation (SPPMP) scheme. We
examine the regime dependence of the impacts with 48 h forecasts of days with weakly and strongly forced convection, as
well as winter storms.
Early in the forecasts, the two stochastic parameterizations have different effects. The PSP scheme responds to boundary
layer turbulence and produces strong perturbations and perturbation growth in phase with the diurnal cycle of convection.
In regions with existing precipitation, SPPMP produces perturbations that grow more slowly with lead time, independent of
the time of day. Perturbation growth in the PSP experiments is stronger for convective weather and can increase the total
precipitation by triggering new convection, but SPPMP can dominate when precipitation occurs under stable conditions at
night or in winter storms. Significantly, the differences between the two schemes are relatively short-lived, and within a day
of simulation, the amplitude and structure of differences introduced by both schemes are similar. This is found to be
associated with saturation of perturbation growth on small scales (up to about 50 km). The locations and amplitudes of
upscale perturbation growth appear to be determined by the larger-scale dynamics, independent of the details of the
stochastic physics.

The oceans and ice components contribute to the seasonal and longer time scales in the earth system. Ocean heat (and carbon) uptake is controlled by a combination of short space-and-time scale processes at the air-sea interface and slower processes connecting to the abyssal ocean which acts as an immense reservoir. Of these processes, mesoscale eddies play a major role in the vertical redistribution of heat within the ocean and we know that the balance of terms in a model heat budget is strongly dependent on how well eddies are resolved (or otherwise represented by parameterizations). We also have demonstrated that the formulation and numerical methods can lead to severe distortion of the heat budget, partly due to spurious numerical mixing. Application of machine learning in ocean modeling needs to meet the same requirements as numerics and physical parameterizations for the resulting heat budget to be trusted. With the transition to routine eddy-permitting or eddy-resolving models, and the advent of hybrid-ML models and ocean emulators, we discuss where we stand with respect to the fidelity of the representation of ocean heat uptake. We’ll take a deep dive into several aspects of the problem, ranging from the choice of numerical methods to the question of parameterization-vs-resolution.

The data-driven revolution has transformed Earth system modelling. The very recently introduced generative diffusion models learn how to produce forecast samples based on given initial conditions and possibly forcings. These generative models outperform not only deterministic data-driven models but also ensemble forecasts with geophysical models. They additionally exhibit physically consistent predictions, previously unseen for completely data-driven models. Based on these developments, we introduce the very first efficient Arctic-wide generative sea-ice model that predicts all important sea-ice variables at the same time in a 12-hour window. Trained on more than 20 years of data from state-of-the-art coupled sea-ice-ocean simulations, the model runs at a 1/4° (around 12 km) resolution for the full Arctic. To make the model more efficient, we have developed a transformer-only architecture that works with relative positional information, and we apply a domain decomposition approach during training and forecasting. The resulting model improves upon baseline methods, like a simple dynamical model and deterministic data-driven models, generates physically plausible trajectories, and can seemingly extrapolate to previously completely unseen conditions. Just trained for a 12-hour forecast, the model remains stable even over years of prediction with a similar seasonal cycle as observed with geophysical models. These results are a very promising step towards physically plausible data-driven modelling for long time scales as needed in climate projections. Furthermore, the very efficient approach allows us to train the model and run the Arctic-wide forecast on a single consumer-grade GPU, reaching around 144 simulation years per day, presenting itself as leap forward in the efficiency of such generative models.

This talk will review the history of meteorological observations and examine how we arrived at the astonishingly comprehensive global observing system that currently serves operational weather forecasting. To understand the evolution of the global observing system, we will look at how requirements have changed from the early days of Numerical Weather Prediction (NWP) to the development of advanced Data Assimilation (DA) systems able to digest many millions of highly heterogenous observations every hour. However, our capability to deploy new observation systems (particularly in space) can operate on very different timescales compared changes in requirements and our ability to effectively exploit some observations. Thus, we (the NWP community in collaboration with data providers) are forced to speculate well ahead of time which observations will be needed and which will have the greatest impact. This talk will demonstrate some of the new tools that have been developed to guide planning of the future observation network. Finally, in this talk we will we look at how we sit on the edge of a revolution (not evolution) in the way observations are used for NWP. Traditionally, the role of observations and DA systems was limited to providing initial conditions for forecast models. New and exciting developments are in play, where observations can now directly influence weather forecasts over the entire forecast range, and forecast models are even being built directly in observations space.

Global Navigation Satellite System (GNSS) is an established atmospheric monitoring technique delivering water vapour data in near-real time with a latency of 90 min for operational Numerical Weather Prediction in Europe within the GNSS water vapour service (E-GVAP). The advancement of GNSS processing made the quality of real-time GNSS tropospheric products comparable to near-real-time solutions. In addition, they can be provided with a temporal resolution of 5 min and latency of 10 min, suitable for severe weather nowcasting. This paper exploits the added value of sub-hourly real-time GNSS tropospheric products for the nowcasting of convective storms in Bulgaria. A convective Storm Demonstrator (Storm Demo) is build using real-time GNSS tropospheric products and Instability Indices to derive site-specific threshold values in support of public weather and hail suppression services. The Storm Demo targets the development of service featuring GNSS products for two regions with hail suppression operations in Bulgaria, where thunderstorms and hail events occur between May and September, with a peak in July. The Storm Demo real-time Precise Point Positioning processing is conducted with the G-Nut software with a temporal resolution of 15 min for 12 ground-based GNSS stations in Bulgaria. Real-time data evaluation is done using reprocessed products and the achieved precision is below 9 mm, which is within the nowcasting requirements of the World Meteorologic Organisation. For the period May–September 2021, the seasonal classification function for thunderstorm nowcasting is computed and evaluated. The probability of thunderstorm detection is 83%, with a false alarm ration of 38%. The added value of the high temporal resolution of the GNSS tropospheric gradients is investigated for a storm case on 24–30 August 2021. Real-time tropospheric products and classification functions are integrated and updated in real-time on a publicly accessible geoportal.

Under the leadership of ECMWF, the EU Copernicus Climate Change Service (C3S) and the EU Copernicus Atmosphere Monitoring Service (CAMS) have developed into highly recognised international providers of climate and environmental information. Their information is an essential component of climate services at the national level. Here we explore the role of climate services in Germany and highlight key challenges and opportunities in ensuring seamless collaboration between national and international climate initiatives. We provide insights into how a national meteorological service operates, their impact on decision-making and their integration within the broader European framework.

As a federal agency, the German Meteorological Service, Deutscher Wetterdienst (DWD), is responsible for providing meteorological and climatological services for the general public, the administration at all governmental levels as well as for manifold economic sectors. These responsibilities fall within the framework of a multitude of national and international climate related regulations. DWD delivers crucial data and consultation to all players in climate change adaptation in Germany. The provision of the best suitable service for each specific application is further supported by the collaboration of DWD with manifold national and international partners. An example at the national level is the DAS core service "climate and water". By exemplifying the key aspects within the frame of these regulations, the role of a national meteorological service is briefly highlighted, challenges portrayed and solutions provided. In addition to this, the role of DWD as the national coordination of the C3S and CAMS will be explored. Special focus will be given to the ‘last mile’ of national climate services – how European data is effectively tailored to and complemented by high-quality national data in order to support accurate national and regional climate monitoring, forecasting, projection, and thereby decision-making in support of climate change adaptation and mitigation.

Extreme precipitation events are expected to become more frequent and more intense due to climate change. The impacts of intense precipitation are often local and downstream river flooding that damages buildings and critical infrastructure and can culminate in fatalities. In 2024, more than 460.000 hectares of land were flooded due to extreme precipitation, affecting more than 1 million people, according to the Copernicus Emergency Management Service (CEMS). While flood early warning systems, such as the European Flood Awareness System (EFAS), contain flash flood and extreme precipitation products, their primary focus remains on river flooding. Consequently, operational warning systems do not yet fully account for the increasing risk of impermeable urban areas to localized extreme precipitation. Although numerical weather prediction models aim to improve forecasting of such events, they continue to exhibit biases, often underestimating the total precipitation during extreme events, leading to misrepresentations of the potential impacts.

Here, we present a novel forecast for extreme precipitation that addresses the gap in existing warning systems and targets the prediction of pluvial flooding over urban areas up to 5 days in advance. To circumvent the use of biased precipitation forecasts, we focus on the prediction of the rarity of an event: how often did the predicted event intensity occur in the past? Exploiting the information embedded in existing numerical ensemble forecasts and hindcasts, we predict the likelihood of extreme precipitation to exceed critical return period thresholds. These forecasts are impact-focused and predict the probability of approaching a one-in-10-years, one-in-20-years, and one-in-50-years event. We present results for extreme precipitation events in 2024 to demonstrate the added value of these impact-centered forecasts. Further, to transform these probabilistic forecasts into actionable information, we estimate the potential risk by mapping the likelihood and potential impacts as a function of lead time. An evaluation over 30 activations from CEMS Rapid Mapping in 2024 demonstrates that the derived risk index for extreme precipitation provides reliable and actionable information that could be used to supplement the existing warning systems. A corresponding operational system is being tested for all medium-to-large cities in Spain, Italy, France, Germany, and Mozambique, using an email notification as part of the Copernicus Evolution service implemented by the HORIZON project CENTAUR.