The Anemoi Framework (EMS Technology Award 2025) represents a pioneering effort in the integration of machine learning (ML) with meteorological forecasting, all developed through a collaborative European initiative. Data-driven machine learning approaches are currently revolutionizing weather prediction with state-of-the-art models outperforming equation-based forecasting system across a wide range of scores. Through this, they have rapidly emerged as candidates for operational weather forecasting systems, delivering accurate forecasts for low computational cost. Designed to enhance the accuracy, efficiency, and accessibility of data-driven weather forecasting, Anemoi builds upon advanced ML techniques and a modular, open-source architecture to democratize access to cutting-edge forecasting tools.

The framework is organised into distinct Python packages covering the entire machine learning lifecycle—from the creation of customised datasets from diverse meteorological sources to the development and training of advanced deep learning graph models. Once a model is trained, Anemoi enables users to run it in operations, using the outputs of physics-based NWP analyses or ensembles as initial conditions, while maintaining comprehensive lineage tracking.

Anemoi has been instrumental in the development of ML-powered weather models including AIFS (ECMWF), Bris (MET Norway), and AICON (DWD), and supports both global and limited area domain models in both deterministic and ensemble settings. These applications demonstrate Anemoi’s potential to enhance forecasting accuracy by integrating ML techniques into existing systems.

More than just a technical framework, Anemoi represents a collaborative effort among meteorological services, researchers, and technologists, fostering knowledge exchange and innovation.

How to cite: Wijnands, J. and Jacob, M. and the Anemoi team: Anemoi: A New Collaborative Framework for Data-driven Weather Forecasting, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-15, https://doi.org/10.5194/ems2025-15, 2025.

Accurate weather prediction in complex topographical regions remains a significant challenge for numerical weather prediction (NWP). While global NWP models have advanced considerably, limited-area models still struggle to capture the fine-scale atmospheric processes critical for regions with complex topography, such as the Alps. Recent AI-based modelling approaches are transforming this landscape, with graph-based AI models showing great potential in capturing complex dynamics and offering flexibility across diverse data modalities. Most of these models, however, are developed for global applications and trained on global datasets like the ERA5 reanalysis. Despite the extensive temporal coverage of ERA5 encompassing diverse weather conditions, its limited spatial resolution is constraining its ability to resolve finer-scale atmospheric processes critical for limited-area modelling.

Building on these developments, we aim to leverage the power of graph-based AI models and the valuable information provided by ERA5 for limited-area modelling. Our approach utilizes the Anemoi framework to train specialized models for the Greater Alpine Region, centred over Austria, using a high-resolution regional reanalysis ensemble dataset. Specifically, we employ the Austrian ReAnalysis (ARA) dataset with its superior 2.5 km spatial resolution and 3-hourly temporal resolution in reanalysis to better capture localized weather phenomena over regions with complex topography regions that global models typically fail to capture accurately.

Our methodology encompasses multiple AI modelling approaches to determine the most effective forecasting strategy. These include global models focusing on our local study domain with a stretched-grid structure and limited-area models forced with external boundary conditions. Furthermore, we conduct transfer learning experiments that harmonise information from the long temporal record of ERA5 with the high spatial resolution of the ARA dataset, creating a more robust prediction system by mitigating issues in model training from the comparatively limited temporal extent of ARA.
To evaluate their effectiveness, we assess the performance of these different modelling approaches, comparing them against traditional physics-based models and validating the results against point-based observations. Additionally, we provide detailed examinations of these findings through case studies of impactful weather events in Austria, highlighting real-world applications of our approach.

The insights from this research will guide the next steps towards harnessing both large-scale and localised information with an AI-based approach, ultimately advancing the accuracy and relevance of limited-area models in operational weather forecasting. 

How to cite: Küçük, Ç., Gfäller, P., Schicker, I., Awan, N. K., and Kann, A.: Advancing Limited-area Numerical Weather Prediction with Graph-based AI: Leveraging Global and Regional Reanalysis Data for the Greater Alpine Region, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-239, https://doi.org/10.5194/ems2025-239, 2025.

The generation of accurate and reliable short-term forecasts (<12 hours) of near-surface (∼ 10m above ground level) wind speed fields, hereinafter called NSWS, are crucial for various socioeconomic and environmental applications. However, monitoring and forecasting NSWS is challenging due to its inherent space-time variability, especially in regions with complex orography such as the Iberian Peninsula in Spain.

Traditional NSWS forecasting methods relies on Numerical Weather Prediction (NWP) models, which require significant computational resources. In addition, these NWP models often yield inaccurate results, especially in regions with complex orography. As a more efficient alternative to this constraint, here we explore the ability to Artificial Intelligence (AI) methods to enhance the efficiency and accuracy of short-term NSWS predictions. We propose the use of two deep learning methods:

1) A U-Net architecture based on Partial Convolutions to generate high-resolution hourly NSWS maps from station-based observations[3].

2) An encoder-decoder architecture based on mixed convolutional and recurrent (ConvLSTM) layers to predict short-term NSWS maps using the generated infilled data as input[4].

This real-time AI-based product, designed as an early warning system, generates high-resolution (3/9 km) short-term (12 h; σ=1 h) NSWS forecasts in near real-time (seconds), achieving high correlation and low prediction errors.

Meteorological stations provide accurate, site-specific wind observations but have limited spatial coverage, especially in mountainous or remote areas. In contrast, reanalysis products offer full coverage at low resolution but fail to accurately reproduce local wind conditions. Our AI-based tool bridges these gaps by combining station and simulation data, though its inference relies solely on station data, making it a cost-effective alternative to NWP models. Observations come from Spanish Meteorological Agency (AEMET)[1], while the reanalysis used is ERA5-Land (9 km)[2].

Beyond performance evaluation, we apply well-established interpretability techniques to analyze the model’s decision-making process:

1. Feature Importance methods were used to evaluate the relevance of each input time step. Both Feature Permutation and Feature Ablation revealed an expected exponential decline in importance over time, but also highlighted that time steps around 7 hours in the past play a key role in accurate forecasting, underscoring the value of long-term information.

2. Pixel Attribution techniques were used to identify important spatial regions in the input wind speed maps. Despite differences, all methods consistently emphasized regions where unexpected wind patterns or extreme events occur, revealing the model’s primary focus. Guided Grad-CAM offered the most interpretable results by combining coarse and fine details.

These interpretability analyses enhance trust in AI-driven forecasts while guiding improvements in model development. While tested on the Iberian Peninsula, the approach is adaptable to other regions. This scalable AI-based method enhances short-term NSWS forecasting for AEMET and other meteorological services, showcasing AI’s potential to improve both forecast accuracy and operational efficiency in meteorology.

How to cite: Martínez-Roig, M., Azorin-Molina, C., P. Plaza, N., Andres.Martin, M., Monsalvez-Pozo, K., Chen, D., Zeng, Z., Vicente-Serrano, S. M., McVicar, T. R., Guijarro, J. A., and Safaei-Pirooz, A. A.: Evaluation and Interpretability of AI-Driven Short-Term Wind Speed Forecasting., EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-494, https://doi.org/10.5194/ems2025-494, 2025.

In recent years, models based on artificial intelligence (AI) have become equally good or even slightly better at predicting the weather as standard operational models, which are based on solving physical equations. Although the grid size of the AI weather models is similar to that of to global operational models, it has been widely noted that forecasts from the AI models are overly smooth. This smoothness poses a potential issue for ensemble generation and for the simulation of extreme weather events, which are often caused by a superposition of multiple spatial scales. In this study, we develop a mathematical argument to better understand the reason for this low "effective resolution" of deterministic AI weather models. We find that an ideal, perfectly trained AI model follows the mean of the predictive distribution for the lead time interval which is used in its loss function during training. We demonstrate the consequences and limitations of this result with forecast data from various AI models, including Aurora, Pangu, GraphCast and GenCast.

We further demonstrate that a low effective resolution leads to better mean-square forecast error scores by reducing the double-penalty effect, especially at longer forecast lead time. To avoid this often unwanted effect, we suggest a spectral rescaling method for a fairer comparison of two models with different effective resolution. By applying this method to the AI forecasts and the ensemble and deterministic forecasts from the European Centre for Medium Range Weather Forecasting (ECMWF) we estimate to what extent the reported advantages of the AI models are only related to their smoothing. Our results will help users of AI forecasts to interpret their output correctly and guide AI developers in the design of loss function and training protocols.

How to cite: Selz, T., Bruinsma, W., Craig, G., Markou, S., Turner, R., and Vaughan, A.: On the effective resolution of AI weather models, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-541, https://doi.org/10.5194/ems2025-541, 2025.

In physics-based numerical weather prediction models, underlying physical laws and numerical discretisation schemes offer some interpretability of the simulated processes and point to potential sources of errors. In contrast, data-driven deep-learning (DL) models lack explicit physics-based interpretation of their predictions. However, the low computational cost and auto-differentiable implementation allow for fast and extensive diagnostics of the origins of forecast errors using well-established explainable AI methods and traditional diagnostics tools. Here, we combine saliency maps and gridpoint relaxation to perform a multivariate regional analysis of the sources of forecast errors in global DL weather forecasting.

We developed ConvCastNet, a DL global weather prediction model based on depth-wise separable convolutional neural networks. The model predicts 6 atmospheric variables at 13 pressure levels and includes 2 additional single-level prognostic and 7 constant input fields. The forecast is computed on a 3-degree lon-lat grid and uses 12-hour autoregressive time stepping to roll out the forecast. ConvCastNet achieves significant success in predicting geopotential at the 500 hPa pressure level, with 8.5 days of useful forecast (based on an anomaly correlation coefficient greater than 0.6), which makes it a suitable tool for diagnostics of the origin of the forecast error.

We use ConvCastNet to systematically nudge subdomains of the forecast fields for 1) planetary boundary layer, 2) stratosphere, and 3) tropics towards a "true" weather state (reanalysis) and monitor the forecast skill improvements beyond selected subdomains. Our results show that an 8-day mid-latitude weather forecast improves significantly with relaxation in the stratosphere, whereas relaxation in the tropics has no significant impact on mid-latitude. This highlights the importance of accurately representing the stratosphere for medium-range weather prediction and the limited impact of tropical variability on mid-latitude forecasts.

Additionally, we investigate the relationship between model error sensitivity to initial conditions and relaxation experiments. By utilising the model's auto-differentiability, we analyse saliency maps, i.e. the gradients of the forecast errors with respect to input fields, to identify overlapping regions of large error sensitivity and high impact of relaxation to the truth. We find the model sensitivity largely consistent with physics-based expectations, with local errors being sensitive to the upstream dynamics and varying sensitivity to different variables and pressure levels. We believe that this combined approach provides valuable heuristics for diagnosing neural model errors and guiding targeted model improvements.

How to cite: Perkan, U., Zaplotnik, Ž., and Skok, G.: Forecast Error Diagnostics in Neural Weather Models Using Gridpoint Relaxation, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-221, https://doi.org/10.5194/ems2025-221, 2025.

Postprocessing of ensemble forecasts with Ensemble Model Output Statistics (EMOS) can reduce systematic errors in the ensemble mean and spread. Classical EMOS finds linear transformations between the original and postprocessed ensemble mean and variance that optimise the Continuous Ranked Probability Score (CRPS). Within the Weather and Climate Science for Service Partnership - India (WCSSP-India) project HEavy Precipitation Forecast Post-processing over India (HEPPI) we have implemented this approach to postprocess daily precipitation forecasts over India for the monsoon seasons 2018-2022 from the NEPS-G forecasting system run at the National Centre for Medium Range Weather Forecasting (NCMRWF) in Noida. In Angus et al. (2024) we have shown that over most of India EMOS improves the CRPS and the prediction of the probability for heavy precipitation, and also reduces over- or underdispersion in the ensemble.

The EMOS approach is not restricted to the classical linear transformations between the original and postprocessed ensemble mean and variance, and to the optimisation of CRPS. It can be made more flexible by using Machine Learning (ML) to find non-linear transformations that optimise the CRPS or other forecast evaluation criteria. Within the WCSSP-India project HEPPI-ML we have explored different versions of ML-EMOS to postprocess NEPS-G daily precipitation forecasts. These include local and non-local Multilayer Perceptrons, and convolutional neural networks. We will present first evaluation results, including a comparison of ML and classical EMOS.

Angus, M., M. Widmann, A. Orr, G.C. Leckebusch, R. Ashrit, and A. Mitra, 2024: A comparison of two statistical postprocessing methods for heavy‐precipitation forecasts over India during the summer monsoon. Quarterly Journal of the Royal Meteorological Society, 150(761), 1865-1883.

How to cite: Widmann, M., Thakur, C., Angus, M., Ng, K., Gonczol, N., Ashrit, R., Orr, A., Leckebusch, G., Geen, R., and Mitra, A.: Classical and Machine Learning EMOS for postprocessing precipitation forecasts over India, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-288, https://doi.org/10.5194/ems2025-288, 2025.

With improving resolution of NWP products and observation data, efficiency and scalability of processing methods are becoming increasingly relevant. Substituting the inputs with their lower-dimensional representations in downstream computations is a common technique to address this issue.

This study examines the Encoder, Decoder, and the Encoder, Decoder, and Discriminator setup, utilizing Single-Head Vision Transformer (SHViT) as a backbone architecture. Scalability with a further increase in input size and/or resolution was considered in the selection. SHViT architecture demonstrates good scaling with larger input sizes, making it a suitable candidate.

We optimize for representation learning tasks on fields of selected near-surface variables from the CERRA dataset from 01/2010 to 12/2019. The CERRA dataset was chosen due to its availability and diversity within the spatial extent. We prepared a set of smaller regions, focusing on areas with prominent orographical features. The test set consists of regions not used during optimization, while the validation set is a temporal split from the training set regions. During the evaluation, the scaling constants are assumed to be known. Orography and land-sea mask are provided as input channels for the Encoder at inference and for the Discriminator during optimization for each region.

The Discriminator is enhanced with an explicit Haar wavelet transform and spectral power transform of each variable field as additional input channels.  The latent representation is a one-dimensional vector with 256 features. We evaluate the spatial distribution of mean absolute error (MAE) and associated standard deviation.

This study is based on work carried out at IBL Software Engineering, Innovation Department as part of research and development of NWP products post-processing framework.

How to cite: Vozár, M.: Representation learning of near-surface atmospheric fields using SHViT modules, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-563, https://doi.org/10.5194/ems2025-563, 2025.

Reconstruction of near-surface wind speed (NSWS; ~10 m above ground level) from local meteorological station measurements remains an open challenge in climate research. Traditional geostatistical interpolation techniques can provide partial solutions, but their reliability is often limited—especially in regions with complex topography, which are common across Spain. These methods are also highly sensitive to the number of available observations.

This work investigates the potential of state-of-the-art Deep Learning (DL) techniques for NSWS reconstruction. In particular, we employ the Climate Reconstruction AI (CRAI) model, an encoder-decoder architecture based on U-Net with partial convolutions, which we train on the Copernicus European Regional Reanalysis (CERRA) dataset, featuring a temporal resolution of three-hourly data and a spatial resolution of 5.5km. This model learns the spatiotemporal patterns of NSWS and is capable of infilling wind fields from grids with missing values.

To apply the model to real-world conditions, we focus on reconstructing daily-averaged wind speed fields from incomplete grids of observational data provided by AEMET (the State Meteorological Agency of Spain). We evaluate several model variants to assess the influence of auxiliary variables such as 2-m air temperature, surface air pressure, 2-m relative humidity and orography. In addition to the baseline U-Net approach, a convolutional attention mechanism is employed to capture complex interdependencies among variables — for example, compound events involving relative humidity, temperature, orography and wind such as the Foehn effect — as well as an LSTM-based recurrent module to leverage temporal information in the reconstruction process.

The resulting models are benchmarked on CERRA data and subsequently applied to reconstruct NSWS fields from AEMET observations across Spain.

How to cite: Monsalvez-Pozo, K., Martinez-Roig, M., Plaza-Martín, N. P., Azorin-Molina, C., and Plésiat, É.: Deep Learning-Based Reconstruction of Near-Surface Wind Speed Fields from Meteorological Observations, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-550, https://doi.org/10.5194/ems2025-550, 2025.

Observations from weather radars are well established for the estimation of precipitation. However these data are unavailable over the ocean and only sparely available over remote regions. Furthermore, in mountains the retrieval of precipitation is severely compromised by the scan geometry. Finally, non-meteorological noise and the saturation of the radar signal by thunderstorm clouds  also pose seroius problems to accuracy. Satellite data can help to overcome these shortcomings. However, the infrared signal provides mainly information about the cloud top temperature and thus no information about rain droplets. However, during day there is a good correlation of the cloud optical thickness and effective radii as e.g. demonstrated in the PhD of Rob Roebeling. Nevertheless in any case the satellite based precipitation needs to be calibrated. Typically, satellite instruments measuring in the micro-wave are used for this purpose. However, in this study, we use ground based data (omrbometer) for the calibration of the satellite based precipitation rate.  We investigate various machine learning methods (e.g. SVM, transformer based GANs and transfer learning) to find the best way to perform recurrent, near-real time calibration updates based on pre-trained information.  In this way, the spatial information of the satellite precipitation can be optimally combined with the high accuracy of the ground based measurements. The presentation will provide an overview about the developed methods and the validation results and the link of the work to the WMO-AINNP project.  Further a outlook will be given about the nowcasting of satellite based precipitation and its integration into the concept of seamless prediction at DWD and the combination with radar nowcasting.

How to cite: Müller, R.: A novel approach for the estimation of precipitation from geosationary satellites., EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-151, https://doi.org/10.5194/ems2025-151, 2025.

Transitioning from experimental research to operational AI/ML applications is a growing priority in weather and climate. At KNMI, a dedicated MLOps team was established, with the objective of developing the infrastructure and best practices for the implementation of AI/ML models. The resulting setup enables the scalable and reliable delivery of high-quality services to society.

This contribution demonstrates the approach taken through a concrete use case: the Collaborative Quantitative Impact Forecasting (C-QIF) for wildfires. Developed at KNMI in collaboration with the Netherlands Institute for Public Safety (NIPV), C-QIF is a data-driven framework that combines meteorological data with historical wildfire records, providing probabilistic quantitative forecasts of the daily number of wildfires across the Netherlands, up to two weeks in advance. This capability plays a crucial role in enabling decisions on resource allocation and public communication during wildfire events, as well as supporting the professional craftsmanship of operational wildfire experts and first responders.

Originally developed in Octave, C-QIF is being reimplemented in Python and adapted for practical use with cloud-based technologies to enhance scalability and reliability, in order to provide real-time operational services. Data pipelines, model execution, and output delivery are being restructured, with the entire infrastructure deployed on Amazon Web Services (AWS). A self-hosted instance of MLflow, an open-source platform for managing machine learning workflows, is integrated to track model development and ensure reproducibility.

Close collaboration between researchers, engineers, and end users is critical throughout the project. The ongoing interaction ensures alignment between scientific goals and operational needs, and enables the establishment of standard practices for testing, deployment and monitoring — key elements for future model development. This contribution highlights how MLOps capabilities are being built cooperatively, and reflects on the lessons learned in operationalising research models, managing heterogeneous data, and leveraging cloud technologies for AI/ML in practice.

How to cite: Alfonsi, A., van Nieuwenhuizen, J. M., Derks, R. A., Trani, L., and de Baar, J. H. S.: MLOps Practices at KNMI: The Collaborative Quantitative Impact Forecasting Use Case, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-156, https://doi.org/10.5194/ems2025-156, 2025.

MeteoSwiss disseminates its extreme weather warnings through various channels, one of the most important being the MeteoSwiss smart phone app. This app is the most widely used platform for informing the Swiss population, with over 4.8 million installations and a daily user base ranging from 700,000 to 1.8 million people. It provides weather forecasts and displays any active severe weather warnings. When enabled, the app also sends push notifications to alert users about such warnings.

In addition, the MeteoSwiss app allows users to submit their own weather observations, known as "Meteo reports." These crowd-sourced reports can include general weather conditions as well as specific impacts such as strong winds, hail, or heatwaves. Users can contribute their observations in three ways: (i) by uploading situational photos, (ii) by selecting predefined keywords to describe the intensity of the event, and (iii) by submitting free-text descriptions.

As MeteoSwiss advances toward impact-based warnings, these user-generated reports play a key role in improving the understanding of how different weather events affect people and infrastructure. By analyzing the relationship between past warnings and reported impacts, MeteoSwiss can refine its algorithms to better predict the consequences of future events based on physical warning parameters. A data driven model essentially predicts a selection of most likely outcomes in terms of user report based on properties of the issued warning. These include keywords describing the impact and draws from submitted photos deemed representative. This improved understanding supports the development of more accurate extreme weather warnings, ultimately enhancing preparedness and response planning.

How to cite: Zamagni, L., Aeberhard, W., Cheng, Y., Mühlhofer, E., and Mahlstein, I.: Leveraging  crowd-sourced impact data with ML to improve severe weather warnings at MeteoSwiss, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-387, https://doi.org/10.5194/ems2025-387, 2025.

How to cite: Briola, E., Stener Hintz, K., and Denby, L.: Real-Time Detection and Characterisation of Trapped Lee Waves: From Deep Learning Model to Operational Deployment in the Iceland-Greenland Region , EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-283, https://doi.org/10.5194/ems2025-283, 2025.

The climatological characteristics of fog and mist at Pula Airport (Croatia) in the northeastern Adriatic were examined, with a focus on applying machine learning techniques to identify large-scale atmospheric patterns associated with their occurrence. Although the published study includes conventional statistical analyses, the primary focus of this presentation is the application of the Growing Neural Gas (GNG) algorithm—an unsupervised neural network model—for clustering synoptic conditions linked to fog and mist formation.

For this study, high-resolution 10-meter wind and mean sea level pressure (MSLP) data from the ERA5 reanalysis (ECMWF) were used to represent synoptic-scale variability over multiple decades. The GNG algorithm was applied to these spatio-temporal fields to identify recurring circulation patterns and characterize their relationship with fog and mist events observed at the airport.

The GNG algorithm constructs a topological map of high-dimensional input data, dynamically adapting its structure by inserting new units in response to data complexity. This adaptability makes it particularly effective for detecting subtle, evolving patterns in atmospheric fields. In this study, GNG successfully identified pressure configurations—particularly quasi-non-gradient conditions—historically associated with fog and mist formation.

A notable result is the observed decline in the frequency of these favorable synoptic patterns, which correlates with a decreasing trend in fog and mist occurrence at the site. This trend appears to be linked to rising sea surface and near-surface air temperatures, reducing the potential for moisture transport from the sea. These findings demonstrate the value of interpretable machine learning techniques in climatological research and provide insight into ongoing changes in low-visibility weather phenomena in coastal regions.

How to cite: Zoldoš, M. and Džoić, T.: Machine Learning Analysis of Fog and Mist Climatology at Pula Airport, Croatia, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-258, https://doi.org/10.5194/ems2025-258, 2025.

Subseasonal weather predictions for the extratropics remain a significant challenge. Although sources of extratropical subseasonal predictability are linked to slowly evolving components of the atmospheric system, such as tropical modes of variability (e.g., the Madden-Julian Oscillation and tropical waves), these sources are not yet fully leveraged due to systematic  errors in numerical weather prediction models. In particular, short-term errors in the tropics can degrade forecast skill in the extratropics on subseasonal timescales. However, the specific tropical regions where such errors most strongly influence extratropical forecast skill remain unclear. Relaxation experiments using NWP models provide a means to identify these key regions, though such experiments are computationally expensive, especially when run in ensemble mode. In this study, we utilize machine learning-based weather prediction models to perform relaxation experiments across various tropical regions on the subseasonal timescale. Probabilistic forecasts are generated using initial conditions from the Ensemble of Data Assimilations (EDA) system of the European Centre for Medium-Range Weather Forecasts (ECMWF). By evaluating reforecasts for a five year period (2020-2024), which lies outside the training period of the models, this study systematically investigates the role of nudging techniques and the influence of tropical variability, particularly the MJO, on subseasonal forecast skill. Additionally, we evaluate the experiments conditioned on the state of the MJO to assess the regional contributions of tropical predictability to forecast improvements in the extratropics. Our results underscore the potential of nudging as both a diagnostic tool and a means to enhance extended-range forecasting skill, especially when combined with emerging machine learning-based prediction approaches.

How to cite: Li, S. and Quinting, J.: Evaluating the Impact of Relaxation Experiments on Subseasonal Forecast Skill Using Machine Learning Weather Models, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-483, https://doi.org/10.5194/ems2025-483, 2025.

This study aims to enhance the prediction performance of PM2.5 concentrations by applying post-processing bias correction to the forecast outputs of the Community Multiscale Air Quality (CMAQ) model, which utilizes meteorological data. The CMAQ model uses meteorological data and various environmental information to generate air quality predictions, but due to the complexity of atmospheric processes, the model often contains systematic errors. To evaluate and improve the accuracy of these predictions, PM2.5 forecasts from CMAQ were compared with ground-based observations from air quality monitoring stations across South Korea.

Bias correction was performed using Integrated Process Rate (IPR) data, which represents the physical and chemical processes that influence air quality within the CMAQ model. This correction was conducted using linear regression and Deep Neural Network (DNN) methods, both of which have shown promise in improving model predictions in other atmospheric studies.

The training data used for the correction covered air quality data from December 2020 to February 2021, a period representing typical wintertime conditions in South Korea. The bias correction was then applied to data from March 2021. This period is particularly significant as it coincides with South Korea's ‘Seasonal Fine Dust Management System,’ which is designed to address high levels of PM2.5 pollution during the winter months. The analysis used these three months of wintertime data to assess how well the bias correction techniques can improve the CMAQ model's accuracy during this critical pollution season.

The results demonstrate that combining the CMAQ model's predictions with IPR data and machine learning-based bias correction techniques significantly improves the prediction accuracy of PM2.5 concentrations. This study illustrates the potential of post-processing the CMAQ model's forecasts with bias correction methods to refine air quality predictions, especially during periods of elevated pollution. The use of advanced AI techniques such as DNN in this context offers a promising tool for improving the reliability and precision of air quality predictions. These improvements are essential for informing public health strategies, air quality management policies, and pollution control measures, particularly during times of high pollution. By improving the accuracy of PM2.5 predictions, this research contributes to more reliable forecasting systems, supporting better decision-making, policy development, and pollution management, which ultimately improves public health outcomes during critical periods of air pollution.

key word : CMAQ, Bias Correction, Machine Learning, PM2.5

How to cite: Lee, S., Song, H.-J., Oh, H.-R., and Kwon, Y.-C.: Machine Learning-Based Bias Correction for Improving PM2.5 Prediction Performance Using the Community Multiscale Air Quality (CMAQ) Model, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-372, https://doi.org/10.5194/ems2025-372, 2025.

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).

How to cite: Jacob, M. and Potthast, R.: EUMETNET's E-AI Programme: Advancing Weather, Climate, and Environmental Applications through Artificial Intelligence (AI) and Machine Learning (ML), EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-526, https://doi.org/10.5194/ems2025-526, 2025.

In recent years, artificial intelligence has been widely used in the field of meteorology. However, it is perhaps less known that AI-based tools have been used for over 30 years in the modeling of pollutant dispersion in the atmosphere.

In the early 1990s, Slovenia and other parts of Europe faced significant air pollution problems caused by SO₂ emissions from coal-fired power plants without desulfurization systems. Due to the highly complex terrain, simple physical Gaussian dispersion models were not suitable for reconstructing the dispersion of SO₂ in the atmosphere, while complex numerical Lagrangian particle dispersion models were still in development.

Therefore, as early as 1992, we modeled air pollution using a Multilayer Perceptron Artificial Neural Network (MLPANN). We developed the world's first comprehensive method for selecting training samples, features, and the topology of the neural network, enabling the model to learn from measured meteorological variables and pollutant concentrations around the power plant to predict SO₂ concentrations at a selected location in advance.

This type of artificial neural network remains the foundation for various derived structures used for machine learning from data even today.

In the following decades, we expanded the use of MLPANN to ozone prediction. Researchers from other Slovenian research groups also worked on this topic. Meanwhile, the use of these and similar tools in pollutant dispersion modelling began to grow on a global scale.

Together with Brazilian researchers, we successfully applied MLPANN for predicting diffuse solar radiation. In this field, we also managed to develop a spatially transferable model, which is not a common capability for artificial neural network-based models.

For the classification of wind fields based on measurements from ground-based meteorological stations, we used a Kohonen neural network.

Later, we added another tool called "Gaussian processes" (which, like MLPANN, is a universal approximator) and the tool "decision trees." We expanded the use to point-based forecasting of basic meteorological variables. In recent years, with high computational capabilities available even without supercomputers, we have been using these tools for surrogate models that can represent or predict pollutant concentration fields in the atmosphere, not only for individual points but for entire areas around pollution sources.

Other Slovenian groups from Slovenian Meteorological Agency, the Department of Meteorology, and the Faculty of Computer Science have, in recent years, used machine learning for global medium-range forecasts of daily averages of meteorological variables and for post-processing weather forecasts from physical models.

In the presentation, we will explain the basic principles of building models based on artificial neural networks in a way that is understandable to laypeople. For experts in the field, we will showcase numerous examples of their application. For the latter group, we hope these examples will inspire others to expand the use of these modern tools even further.

How to cite: Božnar, M. Z., Mlakar, P., and Grašič, B.: Over 30 Years of Artificial Intelligence Use in Meteorology in Slovenia, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-195, https://doi.org/10.5194/ems2025-195, 2025.

How to cite: Aznar Siguan, G., Nerini, D., Tarin Burriel, N., Tay, P., and de Laroussilhe, H. L.: Enabling Efficient MLOps in Weather and Climate Services at MeteoSwiss, EMS Annual Meeting 2025, Ljubljana, Slovenia, 7–12 Sep 2025, EMS2025-722, https://doi.org/10.5194/ems2025-722, 2025.