Some advanced tools for data visualization and statistical experimentation in the field of operational meteorology are presented. But, yes, this is a joke. Or, better, it's only a a pretext to pay homage and remember Anders Persson. He was a great meteorologist and teacher who, from the ECMWF training courses to the countless lectures and conferences held in Italy, has had a profound influence on the authors and many of their colleagues in their understanding of atmospheric physics, operational meteorology and the implications of judgement under uncertainty.

The preparation of forecast information for the needs of society and government in Bulgaria has a history that officially begins in 1956. Over the past 25 years, information from the forecast system of the ECMWF has been actively included in this history. The growth of the capabilities of the forecast unit, using the available forecast information, is possible thanks to the wide range of products offered by ECMWF. Tracking the types of forecasts issued by the National Institute of Meteorology and Hydrology, using the numerical model of ECMWF, reveals the role of the model in operational synoptic work. The mastery of the products offered is also a result of the trainings by ECMWF, visits to the cooperating member-state countries, as well as the UEF forum. We will present a small part of the products used over the last 25-years and those that are currently part of the daily work of forecasters at NIMH.

ECMWF models are widely used by Météo-France, particularly for operational weather forecasting. They are used in several parts of the world for various phenomena, including national or European forecasts and also in overseas territories like the southwest Indian Ocean for tropical cyclones. The aim of this presentation is to give some feedback on IFS’s models by means of several case studies through comparisons with french national models (ARPEGE).

The Comprehensive Nuclear-Test-Ban Treaty Organization (CTBTO) monitors compliance with the Comprehensive Nuclear-Test-Ban Treaty (CTBT), which bans nuclear explosions. To execute its mandate, CTBTO operates a network of seismic, infrasound, hydroacoustic and radionuclide stations . To link radionuclide detections to their possible source locations, CTBTO also operates a community atmospheric transport model (ATM), called FLEXPART. Essentially, it runs backward in time, i.e. from a sample location to regions on Earth where sampled air masses have been located prior to their arrival at a station. A crucial input into FLEXPART is constituted by meteorological fields, which enable the ATM to be driven according to the relevant meteorological information. For over 20 years now, CTBTO has been using ECMWF meteorological fields to drive its ATM.
For the operational backtracking simulations, the input meteorological fields come from deterministic operational ECMWF analyses intertwined with short forecasts. For its future applications, CTBTO is currently implementing an atmospheric transport ensemble modelling system. The aim of this approach is to represent and quantify the contribution of the meteorological uncertainty to the overall uncertainty in the ATM. For that purpose, CTBTO will make an appeal to ECMWF ensemble products, namely ensemble analyses and ensemble forecasts.
In this presentation, we will discuss applications of deterministic ECMWF products, which CTBTO currently uses for its operational ATM simulations and contrast them with the benefits of the ensemble products we plan to use for the upcoming atmospheric transport ensemble modelling system.

Once ECMWF ensemble prediction system had reached operational maturity in the late 1990s, the question arose of how to effectively visualise the enormous volume of information produced. There were two main branches of products: a geographical visualisation, with maps of averages, probabilities of events, or circulation clusters over an area. These displays had the advantage of showing large-scale structures but with the drawback to flatten out the details of the underlying distribution, details that are very important in cases of extreme events, for example. The other way was to display plumes, at a selected point, showing the time series of all ensemble members for the forecast interval. However, this solution was more suitable for continuous variables, such as temperature, while for precipitation, the result was often difficult to interpret. Hence the need to think of something more compact and effective that could convey the essential elements of the distribution and at the same time allowing to show several variables simultaneously. It was at this point the EPS meteogram was set up. The EPSgram, which has been progressively refined over the years, soon become one of the most viewed and successful products on the ECMWF website. The story is not over because. There are ideas to make it even more informative with specific post-processing. In this short contribution we retrace the history, with some anecdotes, and provide some insights for the future.

This study aims to comprehensively evaluate the forecasting capabilities of ECMWF's Artificial Intelligence Forecasting System (AIFS) specifically for severe weather phenomena in Japan, an area characterized by complex topography and diverse meteorological challenges. By focusing on the performance of this AI-based weather forecasting system in predicting high-impact events, we seek to contribute to the ongoing assessment of next-generation meteorological prediction technologies.

Our research will focus on analyzing AIFS performance across multiple high-impact weather events including typhoons, Baiu frontal systems (East Asian rainy season), localized torrential downpours, and heavy snowfall events. Through detailed case studies, we will assess how AIFS predictions compare with conventional numerical weather prediction models currently in operational use.

The methodology involves a two-pronged approach: First, we will conduct quantitative verification of forecast accuracy using standard meteorological metrics. Second, we will evaluate the physical consistency of AIFS outputs by examining whether its predictions adhere to established meteorological principles. This includes analyzing AIFS-generated thermodynamic diagrams to determine if atmospheric thermodynamic structures are realistically represented and meteorologically interpretable.

Key research questions include: (1) Under what synoptic conditions does AIFS demonstrate superior predictive skill? (2) Are there specific weather phenomena where AIFS consistently outperforms or underperforms compared to traditional models? (3) To what extent does AIFS maintain physical consistency in its predictions despite its AI-based architecture?

Preliminary investigations suggest that AIFS shows promising capabilities, but comprehensive verification across diverse weather regimes is essential for operational implementation considerations. Our findings will provide valuable insights into both the potential and limitations of AI-based weather prediction systems for operational meteorology in regions with complex weather patterns.

Accurate hazard prediction is crucial for mitigating the impacts of severe weather events. This study explores the integration of ECMWF models into a flood forecasting system, combining medium-range forecasts with high-resolution nowcasting techniques. By incorporating numerical weather prediction models alongside weather radar data, the system enhances short- and long-term forecasting capabilities. This approach improves real-time decision-making, providing more precise warnings for extreme rainfall and flood events

In recent years, advances in AI-based weather prediction models have led to notable improvements in forecasting the major meteorological variables traditionally simulated by dynamical models. A particularly remarkable enhancement has been observed in the prediction of tropical cyclone (TC) tracks, which are highly affected by the surrounding environmental flows. In the recent prediction of Typhoon Gaemi, which occurred on July 19, 2024 and made landfall in Taiwan, the AI models showed a faster and more accurate track prediction than the dynamical model, demonstrating the necessity of utilizing AI models in the weather forecast. Previous study comparing TC track forecasts from state-of-the-art AI models—Pangu-Weather, FourCastNet, GraphCast, FuXi, and FengWu—suggested that improved representation of the large-scale environmental flow significantly contributes to improved TC track predictions. Compared to conventional dynamical models, AI-based models appear to correct environmental steering flows more rapidly, resulting in earlier improvements in track forecasts. In particular, the improved TC track prediction in ECMWF's AIFS model is likely attributed to a better simulation of TC translation speed. Despite these advances, AI models tend to underestimate the TC intensity and related precipitation. In addition, a comprehensive evaluation of not only track, intensity, and rainfall but also genesis skill is essential. TC genesis in the western North Pacific (WNP) is significantly linked to the variability of the WNP subtropical high and the monsoonal flows. In this study, we assess the 2020 WNP TC genesis performance using AI models—FuXi, GraphCast, and Pangu-Weather—trained on 0.25° ERA5 reanalysis data provided by WeatherBench 2. The year 2020 was historically notable as no TC formed in July, likely due to strong warming over the equatorial Indian Ocean, while TC genesis was active in August and October.

This presentation reflects on the evolution of ECMWF’s approach to data access, from the early days of licensing to the introduction of open data services. Before the establishment of the Advisory Committee for Data Policy (ACDP) in 2001, ECMWF had already begun to shape its catalogues, define access conditions, and develop the principles that would underpin later policies. These early decisions were closely linked to WMO Resolution 40 and helped lay the foundations for more formal data governance.
Over the following two decades, the ACDP guided incremental changes in policy and delivery, as ECMWF responded to emerging user needs, technological change, and growing expectations for broader access. Co-operation agreements played an important role in extending reach while maintaining protections for Member and Co-operating States. The catalogue itself continued to grow, reflecting the evolution of ECMWF products.
More recently, ECMWF has moved towards a tiered open data model, aimed at simplifying access, improving transparency, and reducing costs. This stepwise approach has involved careful consideration of operational and financial implications, including the ongoing need to sustain services and infrastructure at scale.

The weather ecosystem is undergoing a major transformation. Advancements in AI and machine learning, combined with the rise of commercial Earth observation (EO) constellations and cloud-native platforms, are reshaping how weather data is collected, processed, and delivered. Tech companies like Microsoft and Google are now developing their own global models, while startups are building vertically integrated platforms that bypass traditional value chains.

At the same time, as climate volatility increases, demand for timely, localized, and reliable weather insights is rapidly expanding - not only among traditional users, but across nearly every sector from energy and insurance to agriculture and finance.
In this keynote, we’ll look at what this means for the future of weather forecasting: How can public agencies adapt to meet growing user needs? What role will private providers play in delivering operational value? And how do we ensure that trusted, science-driven weather remains a public good in an increasingly commercial and platformized landscape?

The European Centre for Medium-Range Weather Forecasts (ECMWF) serves as the primary source of
prognostic material for operational weather prediction in our national meteorological service. This
presentation evaluates the current utilization of ECMWF forecasts in operational weather prediction
through reliance on various charts, meteograms, indices, and probability-based products, as well as in
numerical modelling through the use of ECMWF GRIB data in high-resolution Limited Area Models
(LAMs).

Current operational workflows extensively utilize ECMWF's diverse product portfolio including
synoptic charts, ensemble meteograms displaying probabilistic information through box-and-whisker
plots, and specialized indices such as the Extreme Forecast Index (EFI) and Shift of Tails (SOT).
Popular products among operational forecasters include mean sea level pressure & 850 hPa wind
speed, 500 hPa geopotential height & 850 hPa temperature, mean sea level pressure & 200 hPa wind,
precipitation & mean sea level pressure, total cloud cover, wind & relative humidity at various levels,
precipitation type, total precipitation, snowfall in the last 6 hours, as well as Multi-parameter EFI (last
24h), EFI for 2 m temperature, EFI for wind speed, EFI for wind gusts, EFI for precipitation, and EFI
for wave height. These products provide essential guidance for identifying anomalous weather patterns
and assessing forecast uncertainty across multiple time scales.

High-resolution numerical modelling capabilities are enhanced through the utilization of ECMWF
GRIB data as lateral boundary conditions for LAM systems. The improved 9 km ensemble resolution
has significantly enhanced boundary condition quality, enabling more sophisticated mesoscale
ensemble predictions at 1-3 km resolution. This resolution upgrade demonstrates particular value for
capturing orographic precipitation and convective processes that global models traditionally struggle to
resolve.

The presentation addresses the implementation of specialized ECMWF software packages for
encoding, decoding, visualizing, and post-processing GRIB and NetCDF files, and various tools for
processing files accessed from the MARS archive in both operational and research applications. These
software solutions facilitate streamlined data processing workflows and enhance the accessibility of
ECMWF products for our national meteorological service needs.

Discussion will include comprehensive verification approaches combining both subjective assessment
by experienced forecasters and quantitative verification against meteorological station observations
throughout Montenegro. The presentation will demonstrate the application of machine learning bias
correction methods utilizing free-atmosphere variables and ECMWF ensemble meteograms to improve
local forecast accuracy. The higher predictability of upper-level parameters such as 500 hPa
geopotential height, 850 hPa temperature, and 200 hPa wind speed through cost-effective
computational runs of AI-based global models with ai-models software package using initial conditions
from ECMWF provides reliable predictors for AI-based post-processing algorithms.

The presentation concludes with an assessment of how these integrated ECMWF products enhance
operational meteorological services’ capabilities for accurate severe weather prediction and decision
support, demonstrating the continued evolution of ECMWF's products in meeting the demanding
requirements of modern meteorological operations.

Keywords: ECMWF ensemble, EFI, SOT, probability charts, meteograms, LAM boundary conditions,
AI models, forecast verification

2025 marks 50 years of ECMWF - and also 50 years of ESA! It’s a shared anniversary that reflects half a century of scientific symbiosis. For decades, ESA satellites have delivered data that help power ECMWF’s forecasts. In return, ECMWF’s model has become a trusted tool for ESA’s calibration, validation, and mission development.

This partnership isn’t just long-standing; it’s essential. Through our co-operation agreement, we’ve built a trusted framework for data exchange, project collaboration, and mutual support. ESA’s Earth observation missions, from ERS to Envisat and the Sentinel series, have shaped ECMWF’s data assimilation capabilities and extended the reach of global forecasting. At the same time, ECMWF’s analysis and reanalysis products have guided ESA’s retrieval science, supported operational cal/val and data quality, helping to define new requirements for future satellites.

Together, we’ve co-evolved. More than 30 ESA missions have provided satellite data for ECMWF’s operational systems. Our shared work now underpins key components of Copernicus and Destination Earth. We are advancing not just our own goals but Europe's collective capacity to observe, understand, and predict the Earth system.

This joint presentation will reflect on 50 years of ESA–ECMWF collaboration: its foundations, milestones, and the value of long-term partnerships in a fast-changing scientific and environmental landscape. It will also look ahead, exploring how we can keep evolving together to meet new challenges in satellite data integration, model fidelity, and user-driven innovation.

The Weather Company (TWCo) helps people and businesses around the world make more informed decisions and take action in the face of weather. As the world’s leading commercial weather provider, about two billion people and businesses across media, advertising, aviation and more rely on Weather Company forecasts everyday.

The Weather Company is a superuser of ECMWF forecasts to drive weather analysis and forecasting products across its business lines. From consumer-focused mobile app-based weather applications to forecasts for nearly 100 global airlines, ECMWF forecasts have played a prominent role in The Weather Company’s forecast production ecosystem for nearly three decades. More recently, artificial intelligence has had a profound impact on the weather enterprise, and TWCo has emerged as a leader in the adoption of AI-based applications via integration of Deep Learning Numerical Weather Prediction (DLNWP) capabilities, employing ECMWF products as a key ingredient in the pivot toward an AI-driven forecasting paradigm.

This presentation will provide a broad overview of The Weather Company’s use of ECMWF products. It will then focus on two particular recent DLNWP-related advancements at TWCo. The first is the integration of ECMWF AIFS into TWCo’s backend consensus weather forecasting system and the improvements to forecast skill that this integration enabled. We will conclude by describing our use of ECMWF products, including ERA5 and AIFS, in our collaboration with NVIDIA to create DLNWP storm-scale ensembles to enable probabilistic decision support.

Convective hazards pose a significant threat to society and are among the most challenging phenomena for forecasting. Increasing spatial and temporal resolution of ECMWF Forecasting system makes it more and more suitable for applications targeting severe convection. European Severe Storm Laboratory (ESSL) has developed additive logistic regression models for large hail and severe convective wind gust using predictors from ERA 5 reanalysis, reports from the European Severe Weather Database (ESWD) and lightning data from the UK Met Office lightning detection network. Applied on ECMWF reforecasts it has been shown that these statistical models exhibit high skill in the medium range. ECMWF and ESSL collaborate in a project to apply these models on ECMWF ensemble (ENS) forecasts up to 240 hour lead times. Probabilistic products for large hail and severe wind gusts from ENS are undergoing tests and further improvements by re-training with more recent data and experimenting with additional predictors. In this presentation the current status of this work will be provided. The results from statistical assessment and application on some real-time severe convective situations will also be shown.

AICON is DWD's novel machine-learning-based global and regional forecasting system. It is built on the Anemoi framework, which was initiated by ECMWF and its member states in 2024 and subsequently won the EMS Technology Achievement Award in 2025. Like ECMWF's operational AIFS suite and MetNorway's Bris system, AICON is becoming another member of the newly forming consortium of Anemoi-based models.

Machine learning (ML) has recently demonstrated significant potential in weather forecasting, capable of matching classical numerical weather prediction (NWP) models. While most global ML forecasting systems rely on the ERA5 reanalysis data, AICON leverages the consistent global and regional reanalysis datasets ICON-DREAM and ICON-FORCE. DREAM and FORCE are based on the 2024 operational ICON NWP model configuration, and outperform ERA5 in certain near-surface metrics.

This presentation outlines AICON's model architecture and addresses the challenges of handling reanalysis data for training. We present results demonstrating AICON's current forecasting capabilities, showing improvements in accuracy and computational efficiency for operational weather forecasting.

ECMWF’s pioneering medium-range ensemble forecasts in 1992 grew out of earlier work on monthly predictions by Tim Palmer and James Murphy at the Met Office, and the two organisations have developed and exploited ensembles in partnership, more or less, ever since. In the 2000s the Met Office developed and implemented its own ensemble system, MOGREPS, with a focus on the shorter range including high resolution regional ensembles, while always using the ECMWF ensemble as a primary tool in the medium range. Early experiments, plus collaboration through the THORPEX programme, demonstrated the benefits of a multi-model ensemble approach and the Met Office now employs a number of multi-model ensemble tools blending outputs from both ECMWF and MOGREPS, and in some cases NOAA in the USA as well. The IMPROVER post-processing system was designed to blend different NWP ensemble sources appropriate to different forecast ranges to give the best possible seamless probabilistic forecasts, and has recently been extended into Week 2 exploiting ECMWF ensemble data – see separate talk by Mylne and Ayliffe.

Ken Mylne has led much of the Met Office ensemble development and exploitation for over 25 years and has attended almost every UEF, and its forerunner, the “EPS Experts Meeting”, since 1999. The ECMWF 50th Anniversary provides an opportunity to reflect on over 30 years of ensemble forecasting to the point where ECMWF has already brought its ensemble up to the same resolution as it’s single-realisation “deterministic” forecast, and the Met Office plans to do the same next year with both global and UK convective scale ensembles. From being an interesting supplement to “operational” deterministic NWP, ensemble forecasts are now central to operational weather forecasting, just as the rise of AI models and ensembles are starting the next revolution. After some reflection on the past, I will offer a few thoughts on the exciting future beyond my own retirement.

Switzerland has released its national drought monitoring and warning system in May 2025. The public web platform www.drought.admin.ch and the operational warning system were developed across three federal agencies in close collaboration with local decision-makers and end-users. Official drought warnings are now issued through the same channels as other natural hazards such as floods. These include www.natural-hazards.ch and the MeteoSwiss mobile app.

A core component of the system is the integration of the full ensemble of sub-seasonal forecasts from the ECMWF. These forecasts are used in two principal streams:

a) as meteorological forcing for hydrological models to generate probabilistic streamflow and soil moisture forecasts, and
b) in the generation of probabilistic forecasts for meteorological drought indices, achieved by seamlessly combining forecast outputs with multi-decadal observational datasets.

Both streams ultimately enter the combined drought indicator, that underlies the drought warning protocol.

To enhance forecast applicability over Switzerland’s complex terrain, raw IFS outputs are post-processed via statistical downscaling and bias correction using a quantile mapping algorithm. This approach yields marginally calibrated forecasts spanning the medium- to sub-seasonal lead time range. Verification analyses indicate that while predictive skill for daily precipitation diminishes rapidly within a few days, spatio-temporally aggregated drought-relevant metrics—such as precipitation anomalies spanning multiple weeks—retain forecast skill at lead times of three to four weeks. This extended predictive capability is crucial for drought preparedness and thus justifies the use of sub-seasonal forecasts within the new warning system.

Unlike other systems, the Swiss drought warning system is unique in its use of both observations and forecasts to provide drought warnings.

The director of the Emilia-Romagna Region will say a few words to welcome guests to Bologna.

.

Co-authors: Olav Ersland, Lars Falk-Petersen, Håvard Homleid Haugen, Magnus Sikora Ingstad, Máté Mile, Thomas Nipen, Even Nordhagen, Aram Farhad Salihi, Ivar Seierstad, Roel Stappers, Paulina Tedesco.

Bris is a global auto-regressive initial-state machine learning model for weather forecasting, developed within the Anemoi framework and enhanced to provide higher spatial resolution over specific regions of interest. In this presentation, we will provide an overview of the model's architecture and capabilities, including the integration of additional high-resolution surface fields, our approach to ensemble forecasting and increasing the temporal resolution to hourly. We will also share the latest verification results, highlighting the model’s performance across various settings. Finally, we will discuss perspectives from end users.

N/A

.

In the past decade, Taiwan has faced the challenges of extreme weather conditions due to climate change, particularly the difficulties in water resource management caused by drought. This has presented a valuable opportunity to enhance cross-sector collaboration between CWA and WRA through the implementation of climate services.
To facilitate the effective climate services for decision making in recent drought events, some customized climate service products have been designed and widely utilized. Sub-seasonal to seasonal rainfall forecast products in high spatial resolution were developed using a statistic downscaling and bias correction method with the ECMWF forecasts, specifically for reservoir catchment areas. Furthermore, a sub-seasonal tropical cyclone (TC) threat potential forecast product with ECMWF forecast was developed. Currently, CWA and WRA regularly collaborate on water situation information to enhance communication across sectors. These products are also integrated to provide users with the latest forecast information, ensures timely decision-making and disaster prevention preparations to be made in time to reduce the likelihood of damage and loss.
The CWA establishes a robust framework for proactive water resources management. Tailoring climate information and facilitating cross-sector communication will enhance resilience in preparation for frequent extreme events.

.

Estonia’s accession to full membership in the European Centre for Medium-Range Weather Forecasts (ECMWF) in 2020 marked a significant milestone, enabling the country to actively engage in strategic initiatives and collaborative projects. One of the first key opportunities was the submission of a proposal to host a new ECMWF facility, a process that fostered extensive national cooperation and institutional coordination. Although the facility was ultimately established in Bonn, the proposal exercise strengthened Estonia’s inter-institutional collaboration and served as a valuable model for future initiatives.
One of the key advantages of Estonia becoming an ECMWF Member State is access to ECMWF’s high-performance computing resources, which has allowed the country to discontinue the use of its outdated local cluster. Critical numerical model calculations are now securely run at ECMWF, with ample resources also available for other Estonian users.
Collaborations with national academic and research institutions, such as the University of Tartu and Tallinn University of Technology, have furthered operational services, including adaptation of ocean models and contributions to Copernicus Marine and Atmosphere Monitoring Services. These partnerships, while not always directly linked to ECMWF, have been inspired and supported by developments within the organization. ECMWF Council meetings have also served as catalysts for exploring emerging areas, such as artificial intelligence and machine learning.

.

Weather forecasts beyond 7-days are available to the public from many providers. The challenge at this forecast range is to provide a suitable level of detail relative to the skill of the underlying numerical weather prediction models and considering the issue of fundamental predictability. Too specific and the forecast will change frequently and significantly in many synoptic situations from forecast run to run, undermining user confidence. Insufficiently detailed and the forecast provides little value to the user beyond a simple climatology.
There is strong evidence that probabilistic forecasts derived from ensembles, ideally enhanced by calibration, can offer significantly improved forecast skill. The IMPROVER system developed by the Met Office produces such forecasts. IMPROVER has recently been upgraded to ingest ECMWF ensemble data and blend it with the Met Office’s MOGREPS ensemble to extend forecasts into week 2. However, the challenge of presenting the forecasts in a suitable form for the medium range, with scientific integrity and effective decision-support, remains.
The Met Office is preparing to extend the forecast range of its public offering. This talk will cover some of the diagnostics that are being constructed to provide the public with a picture of the likely weather in week 2 forecasts. The value of aggregation over spatial and temporal dimensions in reducing uncertainty and changeability will be discussed, as will the choice of diagnostics to include and omit from our initial week 2 forecasts.

The Faroese Meteorological Office (FMO) has access to meteorological data from ECMWF for operational forecasting, research and development. The data from ECMWF is frequently used for forecasting purposes, especially when forecasting beyond two days. However, the resolution of the models is too coarse to explain local patters in the complex terrain of the Faroe Islands making it a challenge to predict the weather on a local scale.
FMO is therefore developing an operational high-resolution local scale model system based on the WRF model using the ECMWF model as background forcing. With a current horizontal resolution of 500 meters, the goal is to get a better understanding of the local weather patterns and develop a forecasting system that can predict local variations between the fjords and mountains that make up the Faroe Islands. The model system and some preliminary results are briefly presented.
Being a small, complex, and isolated archipelago between Scotland and Iceland with good data coverage, the Faroe Islands are viewed as an ideal test-case for NWP modelling in complex terrain and the intention is to establish future international cooperation in this field.

Modeling Storm Surges in the Singapore Straits Based on Forecast Wind Data

Singapore’s low-lying coastline is increasingly vulnerable to storm surge-induced flooding 1, posing significant risks under climate change and rising sea levels. To improve early warning capabilities, we developed a data-driven model that predicts storm surges in the Singapore Straits using forecast wind data.

Five years (2006–2010) of wind forecasts from ECMWF were used as input 2, while sea level anomaly (SLA) observations from Tanjong Pagar station served as the prediction target. Forecasts with 10-day lead times enabled extended warning horizons. After preprocessing, wind signals were transformed into wind speed and direction, and signal processing techniques extracted statistical features. A custom feature selection algorithm identified 32 key features contributing most to storm surge prediction.

The final model, a fully connected neural network, achieved 80% accuracy, 72% precision, 62% recall, and 71% weighted score on 2010 test data. The results demonstrate the strong predictive capability of forecast wind data for storm surge modeling in Singapore, though challenges remain in capturing extreme events due to data imbalance and missing factors.

Future work will focus on integrating additional predictors (e.g. precipitation, atmospheric pressure), exploring advanced machine learning models such as LSTM networks, and generating synthetic data to improve robustness for operational use.

This abstract summarizes our recent work [3]. The full paper is under submission.

Acknowledgement

This work was supported by the National Research Foundation of Singapore and PUB (Singapore’s National Water Agency) under the Coastal Protection and Flood Management Research Programme, conducted at the Coastal Protection and Flood Resilience Institute (CFI) Singapore.

References

[1] “IN FOCUS: With ‘no place to retreat to’, Singapore advances to protect its coastlines - CNA.” Accessed: Aug. 18, 2024. [Online]. Available: https://www.channelnewsasia.com/singapore/rising-sea-levels-low-lying-vulnerable-coastal-protection-long-island-3955651

[2] “Operational archive | ECMWF.” Accessed: Aug. 11, 2024. [Online]. Available: https://www.ecmwf.int/en/forecasts/dataset/operational-archive

[3] Z. F. F. S. Z. Y. T. H. A. X. L. K.-L. G. and P. C. W. Elisa Y.M. Ang, “Predicting Storm Surges in the Singapore Straits with Forecasted Wind Data,” to be submitted to Applied Ocean Research, 2025.

The Eastern Black Sea Region, one of the areas in Turkey that receives the highest levels of precipitation, frequently experiences flash floods, landslides, and inundations, which underscores the critical importance of accurate weather forecasting for this region. On 3 June 2025, a significant meteorological event resulted in heavy rainfall across the provinces of Trabzon, Giresun, Ordu, and Rize. Flash flood warnings issued by the Turkish State Meteorological Service (TSMS) a day in advance of the hazard helped to minimize potential loss of life and property.

In this study, 24-, 48-, and 72-hour forecasts from ECMWF products (IFS, EPS, EFI, and AIFS) were compared with observed total precipitation data collected at meteorological stations. The comparison focused on the spatial distribution of precipitation, rainfall intensity, and timing accuracy.

Overall, this study demonstrates how ECMWF forecasts perform under extreme weather conditions and provides evaluations that can assist meteorologists and local authorities in making timely and effective decisions.

EUMETNET's E-AI programme aims to leverage Artificial Intelligence (AI) and Machine Learning (ML) to enhance weather, climate, and environmental applications. It is a strategic initiative and was setup by the EUMETNET Assembly as a five-year Optional Programme, which started January 2024. The programme combines the forces of European National Meteorological and Hydrological Services (NMHSs) and external partners, including ECMWF and EUMETSAT, to advance in these areas. To achieve its objectives, a strategic reallocation of development resources towards AI/ML-based techniques and capacity building are required.

E-AI is structured around three primary pillars: (a) Data Curation, (b) Analysis, Modelling, and Post-processing, and (c) Products and Services. These pillars are accompanied by Communication and Training activities, which support the general transition towards AI-based technologies. The programme is guided by its Strategic Expert Group, which has conducted comprehensive assessments of the AI/ML landscape to inform the strategic direction of the NMHSs. By promoting collaborative development under a permissive open licence, E-AI fosters widespread adoption, a culture of openness, and synergistic innovation. In line with its guiding principles, the programme welcomes further collaboration with international partners, academia, and industry.

To pursue its targets, the E-AI programme has organised a series of workshops, online tutorials, and established working groups. The workshops were structured around the three primary pillars, featuring both in-person and online events. These included joint workshops with EUMETSAT on data curation in pillar (a), ECMWF Machine Learning Pilot Project workshops in pillar (b), and workshops on Products and Services in pillar (c). The workshops have engaged approximately 200 scientists, while the online tutorials reached an audience of over 400 individuals. The workshops have also identified interest in establishing about a dozen working groups, focusing on specific aspects of AI and ML, particularly in the areas of products and services. We will present updates on the various activities, including the development of E-AI ML-ready datasets, the exploration of multimodal applications combining large language models with meteorological fields, and approaches to Machine Learning Operations (MLOps).

Artificial Intelligence (AI), and in particular Machine Learning (ML), has rapidly transformed Earth System Science (ESS) in only a few years. Fully data-driven AI/ML emulators are now capable of approximating the behavior of traditional physics-based weather and climate models, often improving skill and with significantly reduced computational cost. Hybrid modeling approaches that integrate AI/ML with physical principles are also gaining traction, offering the potential to enhance model predictive accuracy while preserving interpretability. In parallel, ML tools have been developed to detect, forecast, and analyze a wide range of natural hazards and extreme events, including floods, wildfires, and severe storms.
The role of AI in ESS is expanding quickly, but its full potential remains largely untapped. NSF NCAR envisions a future in which AI/ML is deeply integrated into Earth system research—complementing physics-based methods, accelerating scientific discovery and prediction, and becoming broadly accessible to domain scientists across disciplines. This presentation will highlight several community-driven initiatives at NCAR aimed at aligning our computing infrastructure, data systems, software, services, and workforce development with this vision. We welcome input on these efforts and are eager to explore opportunities for collaboration with our European colleagues.

The World Food Programme (WFP) is partnering with hydrometeorological services to strengthen anticipatory action against climate hazards in several countries, including Mozambique, Zimbabwe, and Malawi, by supporting the development of national drought trigger systems that enable anticipatory action based on seasonal forecasts. In these countries, since 2021, the drought trigger systems have been guided by the ECMWF rainfall ensemble forecast, downscaled and bias-corrected to provide more reliable drought alerts for specific areas ahead of drought onset. Developing these systems requires close collaboration with national agencies to define appropriate drought thresholds and triggers, addressing both technical and coordination challenges. This work not only highlights the potential of anticipatory action, when informed by forecasts, to protect livelihoods, reduce humanitarian costs, and support a more timely and dignified response to drought risk, but also strengthens the capacity of meteorological services and governments in these countries to effectively use forecast information for decision-making under uncertainty.

Recent advancements have positioned Large Language Models (LLMs) as transformative tools for scientific research, capable of addressing complex tasks that require reasoning, problem-solving, and decision-making. Their exceptional capabilities suggest their potential as scientific research assistants, but also highlight the need for holistic, rigorous, and domain-specific evaluation to assess effectiveness in real-world scientific applications.

First, this talk motivates and describes the current effort at Argonne National Laboratory to develop a multifaceted methodology for evaluating AI models as scientific Research Assistants (EAIRA). This methodology incorporates four primary classes of evaluations: 1) Multiple Choice Questions to assess factual recall; 2) Open Response to evaluate advanced reasoning and problem-solving skills; 3) Lab-Style Experiments involving detailed analysis of capabilities as research assistants in controlled environments; and 4) Field-Style Experiments to capture researcher-LLM interactions at scale in a wide range of scientific domains and applications. For each of these four classes of evaluation, we develop testing methods (e.g., benchmarks) and tools for manual and automatic QA generation and validation, as well as for collecting and analyzing researcher-LLM interactions.

We will present a selection of tools and generated benchmarks, as well as the early analysis of the largest Field-Style Experiments to date (the 1,000 Scientists AI JAM). These complementary methods enable a comprehensive analysis of LLM strengths and weaknesses with respect to their scientific knowledge, reasoning abilities, and adaptability. Although developed within a subset of scientific domains, the methodology is designed to be generalizable to a wide range of scientific domains.