Traditional weather forecasting relies on large scale numerical simulations that run on high-performance computing systems. These methods require substantial computational resources, involve complex workflows, and generate large volumes of data that often exceed individual user needs. Forecast-in-a-Box leverages advances in data-driven modelling to greatly reduce computational and energy costs while delivering tailored forecast products directly to users. Partly funded from the European Commission’s Destination Earth initiative, it packages the entire forecasting chain into a simple and user-friendly application. Built on the open-source Anemoi¹ and Earthkit² projects, it offers a reproducible and modular environment that integrates data access, model execution, and visualisation. This enables accurate forecasts that can be run locally on user desktops, on premise computing infrastructure, or in the cloud.

The approach is being evaluated through a World Meteorological Organization (WMO) Integrated Processing and Prediction System (WIPPS) pilot project led by the Norwegian Meteorological Institute (MET Norway). In this project, a fully packaged forecasting system based on affordable hardware is provided to the Malawi Department of Climate Change and Meteorological Services (DCCMS). The forecasting system is driven by Forecast-in-a-Box and leverages MET Norway’s Bris³ model (Norwegian word for “light wind), a high-resolution data driven weather forecasting model built using the Anemoi framework. The solution is designed to be largely self-contained, with the only external dependency being the retrieval of ECMWF analysis dataset for forecast initialisation.

¹https://anemoi.readthedocs.io/en/latest/

²https://earthkit.ecmwf.int

³https://lumi-supercomputer.eu/data-driven-weather-forecasting-model/

How to cite: Carton de Wiart, C., Cook, H., Tuma, V., Wong, J., Futsæter, H. A., Østvand, L., Bønes, V., Moe, B., Kristiansen, J., Hawkes, J., Sandu, I., and Quintino, T.: Forecast-in-a-Box: AI weather forecasting, easy to run and simple to deploy, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20013, https://doi.org/10.5194/egusphere-egu26-20013, 2026.

Recent advancements in machine learning based weather prediction (MLWP) present novel opportunities for downstream applications like forecasting of renewable energy production from intermittent sources, like wind and solar. MLWP models guarantee shorter simulation run times and lower computational costs, allowing faster updates of downstream models and greater flexibility in the generation of weather scenarios.

Forecasting renewable energy generation critically depends on available weather forecast data at adequate temporal and spatial resolution. Using MLWP weather data in energy system modelling and forecasting has been limited by the coarse temporal resolution of the current generation of models (e.g. ECMWF’s AIFS Ensemble model runs at 6-hour time steps).

In Europe, power market participants are increasingly exposed to weather forecast inaccuracies. This is due to the combined effect of how the power price is calculated for each price area, and the recent increase in intermittent renewable installed capacities. In detail, power prices are set each day for the following day by balancing supply and demand for each Market Time Unit (MTU), which are now 15 minutes long. It is then massively important to benchmark weather forecasts on a time resolution closer to the power market MTU, to properly assess which period will potentially be oversupplied, or undersupplied from intermittent renewable sources. In this context, the 6-hour time resolution of current MLWP models becomes a significant limiting factor for their usefulness.

Using NVIDIA’s Earth2Studio framework, we demonstrate an efficient, integrated MLWP pipeline combining the [open source] AIFS model with the ModAFNO time interpolation model to provide 1-hourly time-resolution MLWP data. This interpolated data is applied to our intermittent renewable energy production models to assess the interpolation quality compared the uninterpolated AIFS data and the best-in-class numerical weather prediction data provided by ECMWF’s IFS Ensemble forecast.

How to cite: Brenna Schjønberg, H., Parviero, R., Koch, M., and Carpentieri, A.: ML-based time interpolation of AIFS Ensemble for renewable energy forecasting, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11158, https://doi.org/10.5194/egusphere-egu26-11158, 2026.

Minimizing pixel-wise errors in precipitation nowcasting inherently biases models toward smooth predictions, causing failures in resolving extreme convective events. To address this, we propose IMPA-Net, a meteorology-aware framework centered on spectral consistency. The architecture integrates three innovations: a parameter-free Spatial Mixer to encode multi-variate physical interactions (e.g., terrain-wind coupling); an Integrated Multi-scale Predictive Attention (IMPA) module to capture dynamics from Meso-β to Meso-γ scales; and a Meteorology-Aware Dynamic Loss (MAD-Loss) that employs asymmetric penalties to counteract regression-to-the-mean. Experiments demonstrate a 37.3% relative improvement in HSS for severe convection (≥45 dBZ). Crucially, RAPSD analysis confirms that IMPA-Net maintains spectral energy consistency across high-frequency bands, enabling it to successfully simulate the complex "dissipation-initiation" lifecycle that existing baselines fail to capture. These findings validate that integrating domain knowledge advances the physical plausibility of data-driven forecasting.

How to cite: He, G. and Cui, H.: IMPA-Net: Meteorology-Aware Multi-Scale Fusion and Dynamic Loss for Extreme Radar-Based Precipitation Nowcasting, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3224, https://doi.org/10.5194/egusphere-egu26-3224, 2026.

Machine learning-based global weather forecasts often suffer from coarse spatial resolution, limiting their ability to capture fine-scale temperature variability in regions with complex terrain or strong urban–rural gradients. We present SR-Weather, a two-stage deep learning framework that downscales coarse 0.25° forecasts into 1 km air temperature fields. Our model is trained using ERA5 and MODIS-derived temperature data, and leverages high-resolution auxiliary inputs, including elevation, impervious surface fraction, and spatial information–normalized air temperature to enhance spatial fidelity. Applied to 7-day lead forecasts from the FuXi model, SR-Weather consistently outperforms FuXi’s own 1-day lead predictions, indicating strong capabilities in both resolution enhancement and bias correction. The model also exhibits robustness under cloud-contaminated MODIS observations by reconstructing missing temperature values using auxiliary data. While developed and validated over South Korea, SR-Weather is region-agnostic and applicable globally due to the availability of MODIS inputs and minimal reliance on localized data. These results position SR-Weather as a scalable solution for high-resolution, ML-based weather forecasting.

How to cite: Park, H., Park, S., Kang, D., and Kim, J.-H.: SR-Weather: Super-Resolution Machine Learning Weather Forecast for 1-km Air Temperature Prediction, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9075, https://doi.org/10.5194/egusphere-egu26-9075, 2026.

Extreme precipitation poses significant risks to society and infrastructure, highlighting the urgent need for accurate short-term nowcasting. While deep learning models have shown promise in precipitation forecasting, they often lack integration with physical principles, leading to inconsistencies and limited skill in predicting convective evolution. In this study, we introduce RainCast—a novel generative nowcasting framework that synergistically combines deterministic physical modeling with stochastic generative networks to improve the accuracy and physical consistency of extreme rainfall forecasts.

RainCast integrates a deterministic branch based on Neural Ordinary Differential Equations (Neural ODE) to simulate large-scale advective processes and a generative branch built upon a conditional diffusion model to capture fine-scale stochastic variability. The model is guided by key physical features such as flow fields, vorticity, and divergence derived from dual-polarization radar observations, which provide essential dynamical information about convective systems. We train and evaluate the framework using vertically integrated liquid water (VIL) data from dual-polarization radars in China (GD-SPOL) and North America (SEVIR).

Quantitative assessments demonstrate that RainCast significantly outperforms existing nowcasting methods such as SimVP, SwinLSTM, and NowcastNet. On the GD-SPOL dataset, RainCast improves the Critical Success Index (CSI) for intense convection (VIL ≥ 160) by up to 14.1% at 90-minute lead times. Structural similarity metrics also show substantial gains, with reductions in Fréchet Video Distance (FVD) by 25.4% and Learned Perceptual Image Patch Similarity (LPIPS) by 44.6%. Case studies further illustrate RainCast’s ability to realistically simulate the evolution of organized convective systems, including squall lines and multicell storms, while maintaining physical coherence in wind field retrievals.

Our results underscore the value of embedding physical guidance into generative deep learning architectures for convective nowcasting. The RainCast framework represents a meaningful step toward more reliable, interpretable, and physically consistent nowcasting of extreme precipitation, with potential applications in operational meteorology and disaster preparedness.

How to cite: Pan, X. and Zhao, K.: Physics-Guided Generative Nowcasting of Extreme Precipitation with Dual-Polarization Radar and Neural ODE-Diffusion Framework, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11714, https://doi.org/10.5194/egusphere-egu26-11714, 2026.

Thunderstorms pose significant risks to society and the economy due to hazards such as heavy precipitation, hail, and strong winds, which is why accurate forecasts are required to mitigate their impacts. Convection-permitting numerical weather prediction (NWP) models can explicitly resolve convective processes, but predicting thunderstorms from their output remains challenging since there is no obvious state variable that directly indicates thunderstorm occurrence. Instead, many approaches rely on combining multiple convective indices, such as convective available potential energy (CAPE), which are derived from state variables like temperature, pressure, and specific humidity, and act as surrogates for thunderstorms.

In this study, we present a deep neural network model that bypasses surrogate variables and instead directly processes the vertical profiles of state variables provided by convection-permitting forecasts. Our model, SALAMA 1D, analyzes ten different NWP output fields, such as wind velocity, temperature, and ice particle mixing ratios, across the vertical dimension, to produce the corresponding probability of thunderstorm occurrence. The model’s architecture is motivated by physics-based considerations and symmetry principles, combining sparse and dense layers to produce well-calibrated, pointwise probabilities of thunderstorm occurrence, while remaining lightweight.

We trained our model on two summers of forecast data from ICON-D2-EPS, a convection-permitting ensemble weather model for Central Europe operationally run by the German Meteorological Service (DWD), using the lightning detection network LINET as the ground truth for thunderstorm occurrences. Our results demonstrate that, up to lead times of (at least) 11 hours, SALAMA 1D outperforms a comparable machine learning model that relies solely on thunderstorm surrogate variables. Additionally, a sensitivity analysis using saliency maps indicates that the patterns learnt by our model are to a considerable extent physically interpretable. Finally, we show that spatial coverage can be extended to all of Europe by retraining on ICON-EU reanalysis data. Our work advances NWP-based thunderstorm forecasting by demonstrating the potential of deep learning to extract predictive information from high-dimensional NWP data—without sacrificing model interpretability.

How to cite: Vahid Yousefnia, K., Metzl, C., and Bölle, T.: SALAMA 1D: Deep-learning-based identification of thunderstorm occurrence in NWP forecasts without relying on convective indices, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13981, https://doi.org/10.5194/egusphere-egu26-13981, 2026.

Accurate real-time prediction of heavy precipitation is essential for disaster prevention. It remains a challenge for operational meteorology, especially for sudden localized convective storms for which traditional radar and observation extrapolation methods struggle to capture their rapid vertical development, which typically originate at altitudes of 4--8 km before descending to the surface in about 10 minutes.  

In Japan, three Multi-Parameter Phased Array Weather Radars (MP-PAWR) generating 3D data every 30 seconds with high vertical resolution have been deployed. Leveraging these dense 4D observations, an AI-based model produces real-time nowcasts (very short-term forecasts) with high-resolution of 500 m and 10-minute lead time. Updated every 30 seconds, our nowcasts outperform traditional methods for predicting the onset and the dissipation of localized convective precipitation. However, performance is degraded during the mature phase of the storm when its structure becomes more complex (e.g., overlapping  convective cells in different lifecycle states, domination of horizontal motion in radar pattern changes) (Baron et al., 2025a).

Two major improvements are currently being investigated: 1) a Quantile Regression Neural Network (QRNN) technique has been integrated to assess the probability distribution of possible nowcasts and thus provide credible intervals (Baron et al., 2025b), and 2) a better representation of 3D motion is being implemented, as it plays a critical role during the mature phase of storms. The new version of the model will integrate two separate modules: one specialized for capturing 3D-motion vectors, while the second predicts rainfall intensity with motion guidance. Both modules use the current nowcast model architecture which has demonstrated solid performance. The motion module is trained using 3D motion vectors derived directly from the radar observations through a 3D Tracking Radar Echoes by Correlation (TREC) method originally designed for PAWR extrapolation (Otsuka et al., 2016).

This study will present these developments with a special focus on the motion guidance module that is being implemented. The limitations of our approach will also be discussed (e.g., QRNN vs diffusion model, TREC limitation for weak gradient cases, no information on rain precursors and mesoscale scales).

Baron et al., 2025a: “Real-time nowcasting of sudden heavy rainfall using artificial neural network and multi-parameter phased array radar”, SOLA, https://doi.org/10.2151/sola.2025-039

Baron et al., 2025b: “3D Precipitation Nowcasting from Phased Array Radar with Uncertainty Estimation Using a Quantile Regression Neural Network”, IEEE RadarConf25,  10.1109/RadarConf2559087.2025.11204931

Otsuka et al., 2016: Precipitation nowcasting with three-dimensional space–time extrapolation of dense and frequent phased-array weather radar observations. Wea. Forecasting, 31, 329–340.

How to cite: Baron, P., Otsuka, S., Amell, A., Kawamura, S., Satoh, S., and Ushio, T.: AI nowcasting of localized heavy precipitation from fast-scanning radar with probabilistic and 3D motion guided prediction, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15508, https://doi.org/10.5194/egusphere-egu26-15508, 2026.

Satellite brightness temperature (BT) observations contain rich information about the horizontal distributions of cloud and rainfall structures; while radiosonde observations provide high-vertical-resolution measurements of temperature, moisture, and wind in the atmosphere. Beyond their traditional use in assimilation and retrieval, this study demonstrates innovative quantitative uses of BT and radiosonde observations for evaluating high-resolution numerical weather prediction (NWP) simulations of tropical cyclones (TCs) and Southwest Vortices (SWVs).

First, we apply BT observations to document  the structural evolution of TCs and SWVs and to directly compare simulated hydrometeor distributions with satellite-observed cloud and precipitation features. These BT-based diagnostics provide objective constraints on model representation of convective initiation and development as well as the impact of diurnal variability.

Second, a BT-based threat-score (BT-TS) framework is introduced to assess the skill of rainfall forecasts with respect to satellite BT observations instead of rainfall observations traditionally used in TS evaluation. Using microwave humidity-sounder channels, the BT-TS metric performs well for assessing rainfall forecast in regions where precipitation observations are sparse or unavailable. The BT-TS forecast results highlight model deficiencies in timing, extent, and intensity of SWV-induced convective rainfall.

Third, radiosonde profiles are used to investigate lower-tropospheric processes critical for vortex evolution, focusing on planetary boundary layer (PBL) height and vertical variability under different vertical-resolution configurations. Verification with high-vertical-resolution (~5–6 m) profiles from 119 Chinese radiosonde stations during the summers of 2021–23 shows that accurately representing PBL height and lower-tropospheric thermodynamic variability requires approximately doubling  the number of ERA5 vertical levels.

Together, these BT- and radiosonde-based diagnostics provide a comprehensive observational framework for evaluating the structural evolution of TCs and mesoscale SWVs. Future work will leverage these insights to refine cloud microphysics schemes, optimize model vertical-resolution design, and enhance the predictability of convection-permitting NWP systems.

How to cite: Zou, X.: Besides Assimilation and Retrieval: Innovative Quantitative Uses of Satellite Brightness Temperatures and Radiosonde Observations, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1397, https://doi.org/10.5194/egusphere-egu26-1397, 2026.

The development of various AI models in recent years has been very promising; the models’ ability to train from reanalysis datasets and evaluate on various metrics opened the door for a variety of new applications. However, the real stress test of any new model is its operational performance - applying predictions to data that weren't available during the model development and assessing the model’s capabilities for predicting real-world scenarios previously unseen. 

In August 2025, we deployed our first high-resolution AI-based model for Iceland and it has been providing us with continuous predictions since then. Here we evaluate forecast skill against surface observations and benchmark against NWP models from the United Weather Centres (UWC) in Denmark and Iceland and our local operational NWP model for Iceland. We analyze the model’s strengths and weaknesses in predicting various weather events and discuss how these characteristics may influence the future model design.

How to cite: Stanisławska, K. and Rögnvaldsson, Ó.: AI model in the real world - analysis of the operational performance of a high-resolution AI weather model for Iceland, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18305, https://doi.org/10.5194/egusphere-egu26-18305, 2026.

Renewable energy sources have an increasingly pivotal role in global electricity generation, which poses challenges to the accurate and efficient meteorological forecasting (such as solar irradiance and hub-height wind speed). The development of AI large models has significantly shortened the time required for medium-range global weather forecast. However, their outputs typically lack high-temporal-resolution solar irradiance (e.g., provided only at 6-hour intervals or not at all), which cannot be directly applied to renewable energy forecasting.

In this work, we propose a machine learning framework to integrate the output variables from AI large models with high-resolution solar irradiance forecasting. Specifically, we train XGBoost models at 15 sites in eastern China using ERA5 reanalysis variables (2020–2023) as inputs and hourly surface solar irradiance derived from Himawari-8/9 satellite as targets. The trained models are evaluated on a 2024 test set driven by ERA5, achieving an annual mean hourly RMSE of 88.5 W m^-2.

To assess the performance of this approach in medium range forecasting, we use hourly forecasts from the GDAS-driven Pangu Weather Model during January and July 2024 as inputs. Over 20 medium-range forecast tests, our approach (Pangu-ML) yields a day-ahead (24-h lead) RMSE of 62.5 (January) /95.4 (July) W m^-2 and a 10-day lead RMSE of 92.3 (January) /110.1 (July) W m^-2. For comparison, we conduct parallel simulations using the GFS-driven WRF v4.6 model at 9-km resolution over eastern China. The WRF-based irradiance forecasts produce day-ahead and 10-day RMSEs of 78.4 (January) /107.6 (July) W m^-2 and 109.8 (January) /130.3 (July) W m^-2 across the 15 sites, demonstrating that Pangu-ML achieves comparable or even superior accuracy.

In summary, our approach takes advantage of the computational efficiency of AI large meteorological models. It enables rapid generation of solar irradiance forecasts with minimal computational cost, thereby offering a practical pathway for subsequent operational ensemble irradiance forecasting.

How to cite: Yan, M., Zhang, M., Yang, K., Shu, Z., and Shao, C.: Bridging AI Large Meteorological Models and Solar Irradiance Forecasting Through Machine Learning Approaches, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10225, https://doi.org/10.5194/egusphere-egu26-10225, 2026.