At RIKEN, the Japan’s national flagship research institute for all sciences, we have been exploring several attempts to integrate data assimilation (DA) and AI/ML. DA integrates the (usually process-driven) model and data, while AI/ML is purely data driven and is proven to be very powerful in many applications. An example is to integrate data-driven AI/ML-based precipitation nowcasting with process-driven numerical weather prediction (NWP). We developed a nowcasting system based on a convolutional long short term memory (LSTM) which takes several time steps of 2-D precipitation image data to predict future images. NWP with radar DA produces future precipitation images, which can be input to the data-driven LSTM to further improve the predicted images. Another example is to develop ML’ed observation operators for satellite radiances. We obtained an improvement by purely ML’ed observation operators without any information from a physically based radiative transfer model. The third example is to use DA with an ML’ed surrogate model for producing more accurate analyses for further training the ML’ed surrogate model. We found that DA with flow-independent background error covariance could produce more accurate ML’ed surrogate model, but ensemble-based DA resulted in a mixed situation probably because the ensemble forecasts by the ML’ed surrogate model may not produce proper error covariance. We also explored developing a limited-area ML’ed surrogate NWP model in collaboration with IMT-Atlantique. In this presentation, we will share the most recent activities of integrating DA and AI/ML at RIKEN.
How to cite: Miyoshi, T., Otsuka, S., Liang, J., Goodliff, M., Saliou, G., Ouala, S., and Tandeo, P.: RIKEN’s activities to integrate data assimilation and AI/ML, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4099, https://doi.org/10.5194/egusphere-egu25-4099, 2025.
Recently, artificial intelligence (AI) models have upended numerical weather prediction (NWP) by achieving performance comparable to or even surpassing that of physics-based NWP models while also significantly reducing computational costs. However, these AI solutions generally operate with initial conditions produced by NWP data assimilation, which remains costly and can suffer from approximations. We introduce OMG-HD, an end-to-end AI weather forecasting model designed to make predictions directly from observational data, including surface observations, radar, and satellite, thus bypassing the data assimilation step. OMG-HD provides kilometer-scale, 12-hour forecasts across the contiguous United States (CONUS) that exhibit greater skill than the leading operational NWP models. Compared to the High-Resolution Rapid Refresh (HRRR), we achieve a 13-48% improvement in RMSE for 2-meter temperature, 10-meter wind speed, 2-meter specific humidity, and surface pressure. These results demonstrate the feasibility of AI-driven end-to-end approaches for operational weather forecasting free of NWP data, offering a promising step towards faster and more accurate weather forecasts to support weather-dependent decision-making.
How to cite: Dong, H.: OMG-HD: A High-Resolution AI Weather Model for End-to-End Forecasts from Observations, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4604, https://doi.org/10.5194/egusphere-egu25-4604, 2025.
This study investigates the uncertainties of two AI-driven meteorological models, Pangu-Weather and Fuxi, in the forecasts of tropical cyclones (TCs) from perspective of target observations. The conditional nonlinear optimal perturbation (CNOP) method is used to identify the sensitive areas for target observations, and the TCs in the Northwest Pacific and Bay of Bengal (BoB), with the latter being often referred as “BoB storms”, are investigated. The results suggest that the predictability of the “Pangu-Weather” model with respect to the BoB storm tracks is limited within 24 hours, and model error effects dominate the uncertainty of the forecasts after 24 hours; while for the TCs in the Northwest Pacific, the Fuxi model is shown to be strongly sensitive to initial perturbations and provide much accurate sensitive areas for target observations associated with TC track forecasts. These results illustrate the uncertainties of the two AI models and provide a theoretical basis for implementing field campaigns for target observations using AI models.
How to cite: Duan, W.: Uncertainty of AI models in tropical cyclone forecasts: target observation perspective, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-2261, https://doi.org/10.5194/egusphere-egu25-2261, 2025.
Accurate characterization of meteorological conditions across urban regions is crucial for developing equitable energy and climate solutions. However, this task faces significant challenges: observational data often lacks spatial and temporal continuity, while numerical weather prediction models can struggle to localize severe weather events due to substantial latency between observation collection and forecast production. To address these challenges we introduce a multi-modal foundation model that rapidly integrates heterogeneous in situ and satellite data to produce a gap-filled global state. Our results show that the atmospheric structures predicted by our model are consistent with observed phenomena such as liquid and frozen precipitation and convection. Further, we apply our model to produce reanalysis and forecast datasets of solar irradiance for renewable energy applications. We also discuss ongoing work to connect global and local systems through a regional high-resolution foundation model, which is driven by multi-modal observations and the dynamics captured by the global model. This research aims to build predictive understanding of environmental systems and their interactions with built environments by improving our ability to forecast phenomena such as cold and warm fronts, convective weather, and their impacts on health, safety, and energy supply and demand.
How to cite: Vandal, T., Duffy, K., and Nachmany, Y.: A Multi-Modal Observation-Driven Foundation Model for Global Data Assimilation, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-21364, https://doi.org/10.5194/egusphere-egu25-21364, 2025.
Incorporating appropriate physical constraints to data assimilation is of great significance for the assimilation of disastrous weather data assimilation and numerical forecasting. Generally, model constraints are often difficult to describe complex sub-grid physical processes with strong nonlinearity and discontinuity, due to difficulties in developing the tangent linear and adjoint. In 4DVar, simplified physical process schemes are often used instead. With the development of artificial intelligence (AI) technology, complex sub-grid physical processes can be probably considered in variational constraints. On the basis of momentum equation constraints, this study introduces sub-grid boundary layer turbulent friction effects through machine learning (ML) and adds them into momentum equation constraints. Firstly, a deep neural network model is used to train the horizontal momentum tendency simulated by the YSU boundary layer parameterization scheme of WRF model. Secondly, under the Ensemble-Var framework of WRFDA, the momentum tendency of the boundary layer is introduced into the weak constraint of the horizontal momentum equation of variational method. The boundary layer turbulent friction term is implemented by embedding a deep neural network model, and its tangent and adjoint operators are developed to construct a ML-DA scheme. Finally, a physical constraint scheme considering the turbulent friction effect of boundary layer is established for data assimilation. The new assimilation scheme is applied to the radial wind assimilation of coastal radar. Numerical simulation experiments on different typical landing typhoons show that the new scheme better described the boundary layer four-force balance during the data assimilation process. Assimilating the direct-observed wind field, the thermal fields such as pressure and temperature are also improved. The new scheme plays a positive role in typhoon intensity and structure forecasting.
How to cite: Li, X.: Implementing the sub-grid boundary layer turbulent effects into variational constraints trough machine learning and its impact on typhoon assimilation , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-2217, https://doi.org/10.5194/egusphere-egu25-2217, 2025.
Operational numerical weather prediction (NWP) systems consist of three fundamental components: the global observing system for data collection, data assimilation (DA) for generating initial conditions (referred to as analysis), and the forecasting model to predict future weather conditions. While NWP have undergone a quiet revolution, with forecast skills progressively improving over the past few decades, their advancement has slowed due to challenges such as high computational costs and the complexities associated with assimilating an increasing volume of observational data and managing finer spatial grids. Advances in machine learning offer an alternative path towards more efficient and accurate weather forecasts. The rise of machine learning based weather forecasting models has also spurred the development of machine learning based DA models or even purely machine learning based weather forecasting systems. This paper introduces FuXi Weather, an end-to-end machine learning based weather forecasting system. FuXi Weather employs specialized data preprocessing and multi-modal data fusion techniques to integrate information from diverse sources under all-sky conditions, including microwave sounders from 3 polar-orbiting satellites and radio occultation data from Global Navigation Satellite System. Operating on a 6-hourly DA and forecasting cycle, FuXi Weather independently generates robust and accurate 10-day global weather forecasts at a spatial resolution of 0.25°. It surpasses the European Centre for Mediumrange Weather Forecasts (ECMWF) high-resolution forecasts (HRES) in terms of predictability, extending the skillful forecast lead times for several key weather variables such as the geopotential height at 500 hPa from 9.25 days to 9.5 days. The system’s high computational efficiency and robust performance, even with limited observations, demonstrates its potential as a promising alternative to traditional NWP systems.
How to cite: Sun, X.: A data-to-forecast machine learning system for global weather, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7911, https://doi.org/10.5194/egusphere-egu25-7911, 2025.
Forecast errors of numerical weather prediction consist of model errors and the growth of initial condition errors, while the initial condition is often optimized based on short-term forecasts. Thus it is difficult to untangle the initial condition error and model error, but it is essential to infer model errors not just for prediction but also for data assimilation (DA). A hybrid deep learning (DL) and DA method is proposed here, aiming to correct model errors. It uses a convolutional neural network (CNN) to extract characteristics of initial conditions and forecast errors, and then provides estimations for model errors. The CNN-based model error estimation method can consider the model error resulted from inaccurate model parameters, or simultaneously consider the model error and initial condition error. Based on the Lorenz05 model, offline and online experiments demonstrate that the CNN-based model error estimation method can effectively correct model errors resulted from inaccurate model parameters, including the forcing F, coupling coefficient c, and relative scale b. For both online and offline model error estimations, simultaneously considering model errors and initial condition errors are beneficial to infer the model errors, compared to considering model errors only. Moreover, using the observations to verify the forecasts has advantages over using the analyses, to estimate the model errors. Using observations can also achieve a faster convergence of model error estimation with online DA than using analyses.
How to cite: Peng, Z., Lei, L., and Tan, Z.-M.: A hybrid deep learning and data assimilation method for model error estimation, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7652, https://doi.org/10.5194/egusphere-egu25-7652, 2025.
Traditional ensemble forecasting based on numerical weather prediction (NWP) models, is constrained by the need for massive computational resources, resulting in limited ensemble sizes. Although emerging artificial intelligence (AI)-based weather models offer high forecast accuracy and improved computational efficiency, they still face considerable challenges in ensemble forecasting applications.
In this study, we propose a fast, physics-constrained perturbation scheme through self-evolution dynamics of AI-based weather model for ensemble forecasting of tropical cyclones (TCs). These initial perturbations are conditioned on specific amplitude and spatial characteristics, exhibiting physically reasonable dynamical growth and spatial covariance. Based on this perturbation scheme, the TC track ensemble forecasts within the AI-based model significantly outperform those from the European Centre for Medium-Range Weather Forecasts (ECMWF) in both deterministic and probability metrics. Notably, we conduct TC track forecasts with 2000 members for the first time, achieving further enhanced forecast skill in probability distribution and extreme scenario of TC movement.
How to cite: Pu, J., Mu, M., Feng, J., Zhong, X., and Li, H.: A Fast Physics-based Perturbation Generator of Machine Learning Weather Model for Efficient Ensemble Forecasts of Tropical Cyclone Track, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-2160, https://doi.org/10.5194/egusphere-egu25-2160, 2025.
In recent years, AI-based weather forecasting models have matched or even outperformed numerical weather prediction systems. However, most of these models have been trained and evaluated on reanalysis datasets like ERA5. These datasets, being products of numerical models, often diverge substantially from actual observations in some crucial variables like near-surface temperature, wind, precipitation and clouds - parameters that hold significant public interest. To address this divergence, we introduce WeatherReal, a novel benchmark dataset for weather forecasting, derived from global near-surface in-situ observations. WeatherReal also features a publicly accessible quality control and evaluation framework. This paper details the sources and processing methodologies underlying the dataset, and further illustrates the advantage of in-situ observations in capturing hyper-local and extreme weather through comparative analyses and case studies. Using WeatherReal, we evaluated several data-driven models and compared them with leading numerical models. Our work aims to advance the AI-based weather forecasting research towards a more application-focused and operationready approach.
How to cite: Jin, W., Weyn, J., and Dong, H.: WeatherReal: A Benchmark Based on In-Situ Observations for Evaluating Weather Models, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7596, https://doi.org/10.5194/egusphere-egu25-7596, 2025.
One of the most critical tools to mitigate the impact of urban flash floods is having an effective and timely early-warning system. The Argentine National Meteorological Service (SMN) is actively working in this direction through the PREVENIR Argentina-Japan cooperation project, which aims to develop an impact-based early-warning and emergency management system for urban flash floods in two Argentine target basins by 2027. As the current SMN operational system consists of 4-km resolution deterministic and warm-start probabilistic forecasts, to provide a more accurate and timely precipitation forecast, under PREVENIR, we are developing a higher-resolution (2-km), rapid-update data assimilation and numerical weather forecasting system coupling the Local Ensemble Transform Kalman Filter (LETKF) with the Weather Research and Forecasting (WRF) model. The system ingests local data from automated surface weather stations and C-band Doppler weather radars to obtain a 40-member analysis ensemble every 5 minutes, and 10-h 20-member extended forecasts every 30 minutes. This work aims to evaluate the performance of the WRF-LETKF prototype system based on a 4-day case study of almost continuous precipitation over the Buenos Aires region in March 2024, which led to urban floods in one of the pilot basins. A preliminary comparison with Radar Quantitative Precipitation Estimation (RQPE) indicates a good performance of the precipitation forecasts and added value for early warning and decision-making.
How to cite: Maldonado, P., Amemiya, A., Dillon, M. E., Gacitua Gutierrez, J., Cutraro, F., Casaretto, G., Ruiz, J., Pulido, M., Garcia Skabar, Y., and Miyoshi, T.: The PREVENIR rapid-update data assimilation and short-range numerical weather prediction system prototype: an urban flood case study over Buenos Aires. , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-12662, https://doi.org/10.5194/egusphere-egu25-12662, 2025.
The operational Tropical Regional Atmospheric Model System (TRAMS) model often underestimates the initial typhoon intensity when using the global analysis field as the initial condition. The tropical cyclone (TC) initialization scheme developed based on incremental analysis updates (IAU) technique can help reduce the initial bias. However, the IAU-based TC initialization scheme only adjusted the wind field at the analysis moment, with other variables adjusted implicitly under the constraints of the model according to the gradually inserted wind increment (named “univariate adjustment scheme” hereafter). The univariate adjustment scheme required approximately 3 to reach a dynamic equilibrium state, limiting the use of hourly TC observation information and dissipating too much meaningful short-wave information of the adjustment increment. To reduce the equilibrium adjustment time, this study constructed a multivariate adjustment IAU-based TC initialization scheme by introducing the gradient wind balance and hydrostatic balance as large-scale constraints. The case sensitivity tests of TC Hato (1713) demonstrated that the multivariate adjustment scheme can reduce the IAU relaxation time to 1 and slightly improve TC forecasts. By incorporating the equilibrium assumption as a strong constraint for the TC axisymmetric component, the multivariate adjustment scheme achieved a faster convergence of the model to its equilibrium state, reducing the loss of useful observed information. This conclusion was further confirmed with 12 other TCs. The IAU-based multivariate adjustment initialization scheme developed in this study provides a foundation for 4-D initialization with hourly TC observations.
How to cite: Zhang, S., Xu, D., Tian, W., and Zhang, B.: Multivariate adjustment in the IAU-based tropical cyclone initialization scheme in TRAMS model, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-3579, https://doi.org/10.5194/egusphere-egu25-3579, 2025.
The occurrence of severe weather events has shown increasing frequency and intensity due to global climate change. The Korean Peninsula, characterized by complex inland orography and surrounded by seas on three sides, exhibits diverse meteorological phenomena significantly influenced by seasonal wind regimes. While forecasters traditionally analyze synoptic conditions using locally available observations, the irregular spatial and temporal distribution of these observations limits their ability to conduct comprehensive three-dimensional analyses. Furthermore, although global model analysis fields have been widely used for operational forecasting, their coarse spatial and temporal resolutions constrain the real-time analysis of localized severe weather events. To address these limitations and enhance nowcasting capabilities, the Korea Meteorological Administration (KMA) has implemented the Korea Analysis System (KAS) since May 2024, a real-time analysis system that utilizes a high-resolution regional model to provide rapid updates of current atmospheric conditions essential for monitoring and predicting mesoscale weather phenomena.
This study evaluates KAS's effectiveness in reproducing real-time atmospheric phenomena and its practical utility for severe weather analysis through extensive synoptic case studies. KAS generates hourly three-dimensional nowcast analysis fields at 3 km resolution by integrating 15 categories of synoptic and non-synoptic observational data with the operational Korean Integrated Model-regional (KIM-regional) forecast fields, assimilating observations up to 15 minutes past each hour to provide near real-time atmospheric conditions. The system demonstrated remarkable capability in capturing critical meteorological features across various weather regimes. During summer, KAS effectively identified precursors of convective precipitation by analyzing real-time low-level convergence zones, dewpoint depression fields, high equivalent potential temperature areas, and vertical p-velocity distributions. The system's skew T-log P diagrams revealed significant Convective Available Potential Energy (CAPE) values, providing quantitative measures of atmospheric instability and potential for convective cloud development and subsequent precipitation in specific regions. In winter scenarios, KAS accurately depicted strong wind variations, including northwesterly cold air flows and easterly winds associated with orographic precipitation. Notably, the system's thermal advection analysis fields effectively identified regions of warm air advection and their interaction with cold air masses, providing crucial indicators for potential snowfall accumulation zones, particularly in areas where warm maritime air masses encounter pre-existing cold air.
The results validate KAS's capability to provide forecasters with coherent three-dimensional nowcast analyses, overcoming the limitations of traditional forecasting methods based on irregularly distributed observations and coarse-resolution global model analyses. This advancement establishes a foundation for improved real-time severe weather detection and forecast accuracy across the Korean Peninsula and East Asia region.
Acknowledgement: This work was supported by Development of Numerical Weather Prediction and Data Application Techniques (KMA2018-00721).
How to cite: Kim, M., Lee, E., Kang, Y., and Lee, Y.: A case study on synoptic analysis using the Korea Analysis System (KAS) to enhance severe weather monitoring over the Korean Peninsula, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-5510, https://doi.org/10.5194/egusphere-egu25-5510, 2025.
Cold waves are one of the most frequent extreme weather events in the winter mid- and high-latitude regions of the Northern Hemisphere. Due to their sudden onset, persistence, and wide-ranging effects, they often cause significant economic and social losses. Although progress has been made in short-term cold wave forecasting, sub-seasonal (3-4 weeks) forecasting remains a major challenge due to the loss of initial condition information, and the complex and nonlinear external forcings. Current numerical models, which dominate operational cold wave forecasting, are computationally expensive and difficult to run large-scale simulations. In contrast, machine learning models, particularly FuXi-S2S developed by Fudan University, offer significant potential due to their efficiency and accuracy in sub-seasonal forecasting.
In the preliminary work, the researcher identified the spatiotemporal characteristics and circulation evolution of cold events in Eurasia, proposing the "Cold Arctic-Warm Continent" mode and its interaction with tropical Pacific signals. Despite improvements in understanding the mechanisms of cold waves, predicting their occurrence at the sub-seasonal scale remains difficult due to uncertainties and complex nonlinear processes. Therefore, exploring new machine learning-based forecasting methods is essential to improve prediction accuracy.
The goal of this study is to: 1) identify key pre-cold wave factors at the sub-seasonal scale in China; 2) develop a probabilistic forecasting scheme based on FuXi-S2S with physically constrained perturbations. The research methodology includes composite analysis and the design of initial perturbations for the FuXi-S2S model with physical constraints, aimed at improving forecast accuracy. By comparing ensemble and deterministic forecasts, this study will evaluate the effectiveness of the proposed scheme and contribute to early warning strategies for cold wave events.
How to cite: Liu, Q.: Research on Subseasonal Ensemble Forecasting of Cold Surges over China Based on Physically-Constrained Perturbations in AI-based Weather Prediction Models, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7063, https://doi.org/10.5194/egusphere-egu25-7063, 2025.
El Niño-Southern Oscillation (ENSO) is the dominant atmosphere-ocean coupled mode of year-to-year variations in the tropical Pacific. It shows diverse spatiotemporal characteristics and casts major influences on seasonal predictions of global weather-climate extrema. Despite numerous dynamical and statistical models for ENSO prediction and predictability studies, they are commonly subjected to one-to-three issues among less skillful simulation of El Niño diversity, huge requirements of computational resources and a low robustness in statistics. Here, an efficient deep-learning model involving nonlinear coupling of multiple variables is independently developed to study the predictability of two types of El Niño events related to initial uncertainty, which is the first kind of predictability problem. The model can skillfully simulate statistically robust features of observed El Niño diversity in terms of periodicity, amplitude, and seasonal phase-locking. Using this model, we have revealed mathematically several new types of fastest-growing initial errors in two types of El Niño predictions based on a novel concept of conditional nonlinear optimal perturbation (CNOP), especially including one that can strengthen central Pacific types of events, which is rarely investigated before. Moreover, CNOPs are superimposed into a numerical model, GFDL CM2p1, for comprehensive validation and growth mechanism mining, which demonstrates the consistent dynamical evolutions of initial errors in both numerical and AI models. Our study represents the first attempt to explore the first kind of ENSO predictability problem from perspectives of nonlinear error evolving dynamics using a data-driven model. This is of great importance as it offers us sufficient confidence to perform ENSO-related (such as the Madden-Julian Oscillation, etc.) mechanisms and predictability studies for future data assimilations and observation programs without strongly relying on dynamical numerical models.
How to cite: Qin, B. and Mu, M.: The First Kind of Predictability Problem of El Niño Predictions in a Multivariate Coupled Data-driven Model, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14917, https://doi.org/10.5194/egusphere-egu25-14917, 2025.
Data assimilation (DA) is the best tool for assessing the state of time-evolving chaotic systems. An estimate (analysis) is derived by combining Information from the most recent and previous batches of observations, the latter of which are carried forward by first guess forecasts started from previous analyses. Error variance in successful DA cycles fluctuate around an expected value. What factors determine this value?
Three parameters are found to describe the level of analysis error variance ( ), or the amount of Information in state estimates: the level of Information extracted from a recent set of observations by a DA system at analysis time ( ), the growth rate of error in the first guess ( ), and the relative weight used for combining Information from the latest set of observations and the first guess ( ). A key recognition of this study is that in DA systems with stationary performance, the gain of Information from the most recent observations, and the loss of Information due to chaotic error growth, in an expected sense, must be equal.
Exploiting this equilibrium relationship, error variance or Information in a state estimate can be expressed as a function of the three driving parameters. Analysis Information linearly and exponentially depend on and alpha, respectively, while the optimal weight is found to be a simple function of the rate of error growth. An evaluation of four operational DA systems reveals that their quality is driven by the amount of observational Information they each extract from the a virtually common set of globally available observations. The ECMWF analysis performs best, with two and a half times lower error variance than in any other system. A simple global adjustment of the relative weight between observations and the first guess may yield an 11-43% reduction in error variance.
How to cite: Feng, J.: An Equilibrium Between the Observational Gain and Chaotic Loss of Information, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-20424, https://doi.org/10.5194/egusphere-egu25-20424, 2025.
The Met Office in the UK is exploring the use of ensemble sensitivity analysis (ESA) as an operational tool to support its upcoming focus on its global ensemble system. Ensemble sensitivity analysis is a technique that identifies atmospheric flow features throughout a forecast period that are relevant to high-impact forecast aspects such as high winds, heavy precipitation, and extreme temperatures (known as response functions). ESA typically highlights the importance of the position or magnitude of features like upper-level troughs, ridges, and wind maxima/minima in the jet stream, as well as structure in low-level pressure and moisture fields, to the response function. Since ESA also identifies specifically how features are associated with differences in high-impact response functions (e.g., an eastward shift of a 300hPa geopotential height trough off the U.S. east coast might be associated with heavier precipitation two days later in the UK), it can add substantial value to the forecasting process through forecaster awareness. This value can be realized through both improved dynamical understanding of high-impact flows and ensemble subsetting, a method that weights ensemble members more if they are more skillful in sensitive areas.
The Met Office in the UK has created a real-time ESA tool for initial evaluation to understand its value in the forecasting process. Wind, precipitation, temperature, and visibility response functions to seven-day forecast time over the UK, both coverage and maximum values, serve as the response functions. Sensitivities to geopotential height and wind speed aloft, surface pressure, and simulated water vapor imagery are produced every six hours from the response function backward to initial forecast time. This presentation involves what operational forecasters and research personnel have learned from day-to-day ensemble sensitivity fields, the use of ESA in the forecasting process, and the climatological nature of sensitivity. Future plans for the Met Office in the UK ESA tool will also be discussed.
How to cite: Ancell, B., Willington, S., Titley, H., Jones, C., Walker, B., Semple, A., Hicks, R., Relton, P., Barciela, R., Etheridge, D., and Roberts, N.: Ensemble Sensitivity Analysis in the Operational Met Office in the UK Ensemble System, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-5156, https://doi.org/10.5194/egusphere-egu25-5156, 2025.
Dual-polarimetric (dual-pol) radar variables, such as differential reflectivity (ZDR) and specific differential phase (KDP), provide valuable information about hydrometeor types, sizes, and water content. A dual-pol radar operator that applies scattering calculations using the T-matrix method for rain and the Rayleigh scattering approximation for snow and graupel can more accurately translate model variables into observed variables. Assimilating dual-pol radar variables in numerical weather prediction models enhances the forecast accuracy for evolving mesoscale precipitation events. Therefore, developing advanced radar observation operators capable of calculating dual-pol radar variables using microphysical variables is crucial.
In this study, an improved observation operator (K-DROP; KNU dual-pol radar observation operator) is developed. The K-DROP restricts the distribution of mixed-phase hydrometeors in regions with strong vertical motions, thereby reducing overestimation of radar variables near the melting layer. Additionally, by incorporating the observed snow axis ratios for cold rain process, the calculation of as a constant value in subfreezing regions is corrected. Observed maximum hydrometeor radius data are also applied, reducing overestimations of and in warm regions. Experiments using LETKF are conducted for both convective and stratiform precipitation cases and compared with the previous observation operator without modifications. While the previous operator improved forecast accuracy compared to control experiments without DA, it showed limited improvements near the melting layer due to reduced hydrometeor mixing ratios and increased downdrafts. In contrast, K-DROP produced more realistic radar variables, stronger updrafts, and higher correlations with observations. These improvements are particularly effective for convective precipitation with localized heavy rainfall, demonstrating the importance of assimilating dual-pol radar variables containing water content information.
Key words: Dual-polarization radar operator, Radar data assimilation, Observation operator, Precipitation forecasting.
Acknowledgments: This work was supported by the National Research Foundation (NRF) grant funded by the Korea government (MSIT) (No. 2022R1A2C1012361), the Korea Meteorological Administration Research and Development Program under Grant RS-2023-00237740 and the Brain Korea 21 program.
How to cite: Lee, J.-W., Min, K.-H., and Lee, G.: Improved Model Prediction with Dual-polarimetric Radar Operator in Ensemble Data Assimilation, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14792, https://doi.org/10.5194/egusphere-egu25-14792, 2025.
In this paper, we explore the prevalent issue of underestimation of extreme precipitation values in deep learning models utilized for precipitation forecasting. We emphasize that this challenge arises from the double penalty phenomenon, which is exacerbated by the joint effect of the commonly adopted mean squared error (MSE) loss function and the intrinsic uncertainty of forecasting tasks. Drawing inspiration from probability-matching ensemble forecasting, we introduce Sort Loss, a straightforward yet highly effective deep learning loss function. By leveraging the ordinal relationships within meteorological data, Sort Loss circumvents the positional information-related double penalty problem. Experimental results from precipitation nowcasting and short-term forecasting tasks demonstrate that Sort Loss effectively diminishes the distributional discrepancies between model forecasts and actual observations. Consequently, it significantly enhances forecasting performance in heavy rainfall scenarios, while simultaneously maintaining stability across other weather conditions. This study offers a novel perspective on optimizing deep learning models for weather forecasting and showcases the potential of applying Sort Loss to improve the accuracy of extreme weather predictions.
How to cite: Cao, Y., Chen, L., and Feng, J.: Imroving Extreme Precipitation Prediction Accuracy: A Novel Probability-Matching-Based Loss Function for Deep Learning Models, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15517, https://doi.org/10.5194/egusphere-egu25-15517, 2025.
Accurate and reliable precipitation nowcasting plays a critical role in disaster prevention and mitigation. The heavy precipitation forecast is a challenging task for most deep learning (DL)-based models. To address this challenge, we develop a novel DL architecture called “multi-scale feature fusion” (MFF) that can give skillful precipitation forecast with a lead time of up to 3 h. The MFF model uses convolution kernels with varying sizes to create multi-scale receptive fields. This helps to capture the movement features of precipitation systems, such as their shape, movement direction, and speed. Additionally, the architecture makes use of the mechanism of discrete probability to reduce uncertainties and forecast errors, enabling it to predict heavy precipitation even at longer lead times. Four-year radar observation data from 2018 to 2021 are used for model training, and the data of 2022 for model testing. The MFF model is compared with three existing extrapolative models: time series residual convolution (TSRC), optical flow (OF), and UNet. The results show that MFF achieves superior forecast skills with high probability of detection (POD), low false alarm rate (FAR), small mean absolute error (MAE), and high structural similarity index (SSIM). Particularly, MFF can predict high-intensity precipitation fields at 3 h lead time, while the other three models cannot. Furthermore, MFF shows improvement in the smoothing effect of the forecast field, as observed from the results of radially averaged power spectral (RAPS).
How to cite: Chen, S., Huang, Q., and Tan, J.: Skillful Precipitation Nowcasting Based on Multi-scale fusion and Radar Observations, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14682, https://doi.org/10.5194/egusphere-egu25-14682, 2025.
Data assimilation (DA) is a statistical approach used to estimate the states of physical systems by integrating prior model predictions (background states xb) with observational data (y). This integration produces an accurate estimate, called analysis states (xa), by sampling or maximizing the posterior likelihood p(xxb, y). In weather forecasting, background states are generated by imperfect models, and the likelihood p(xxb) is often unknown. Observations, sourced from diverse instruments, are mapped to model space using observation operators (H). Effective DA algorithms must accurately estimate p(xxb) while accommodating various observation operators, including those involving sparse, noisy or irregular data.
Traditional DA methods, such as variational assimilation, assume that the background error (x - xb) follows a Gaussian distribution independent of xb. This allows explicit computation of p(xxb, y) and optimization via techniques like gradient descent. While robust to various observation operators, these methods depend heavily on expert knowledge to construct error correlations and are limited by their Gaussian assumptions.
Generative neural networks, particularly diffusion models, have emerged as alternatives for modeling p(xxb). Notable examples include SDA and DiffDA, which use diffusion models to learn background distributions. SDA incorporates observations via diffusion posterior sampling, while DiffDA employs the repaint technique. These approaches improve on traditional methods by capturing more complex distributions but often struggle with sparse, irregular observations. For instance, DiffDA assumes grid-aligned data, while SDA relies on assumptions that can reduce accuracy in real-world scenarios.
In this research, we aim to develop a neural network-based data assimilation algorithm that not only captures the non-Gaussian characteristics of the conditional background distribution for enhanced accuracy but also effectively assimilates data under real-world observations (sparse, noisy and outside of the grid). We introduce VAE-Var, a novel data assimilation algorithm in which a variational autoencoder is first employed to learn the conditional background distribution and then the decoder component is utilized to construct a variational cost function, which, when optimized, yields the analysis states.
Key advantages of VAE-Var include:
- This algorithm inherits the framework of traditional variational assimilation by explicitly modeling the posterior probability function p(xxb, y) and maximizing it to derive the analysis states. As a result, compared to other neural network data assimilation methods such as SDA and DiffDA, VAE-Var can better handle different types of observation operators, particularly irregular observations that do not fall on the grid points of the physical field.
- Unlike traditional variational assimilation algorithms, VAE-Var alleviates the dependence on expert knowledge for constructing the conditional background distribution, enabling the model to effectively capture non-Gaussian structures. This makes VAE-Var perform better in sparse observational settings.
Experiments with the FengWu weather forecasting system at 0.25° resolution show that VAE-Var achieves higher accuracy than DiffDA and traditional algorithms (interpolation and 3DVar) in sparse observational settings. When integrated with FengWu, VAE-Var reliably assimilates real-world GDAS prepbufr observations over a one-year period.
How to cite: Xiao, Y., Jia, Q., Chen, K., Xue, W., and Bai, L.: Variational Autoencoder-Enhanced Variational Methods for Data Assimilation, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14520, https://doi.org/10.5194/egusphere-egu25-14520, 2025.
Posters virtual: Tue, 29 Apr, 14:00–15:45 | vPoster spot 5
EGU25-17568 | ECS | Posters virtual | VPS2
Improving forecasts of extreme precipitation with MAD-WRF mesoscale modelTue, 29 Apr, 14:00–15:45 (CEST) | vP5.4
The Multi-sensor Advection Diffusion Weather Research and Forecast (MAD-WRF) model is a state-of-the-art addition to the WRF model that includes a fast cloud-initialization procedure, making it more suitable for hydrometeors analysis and clouds forecasts. The MAD-WRF cloud initialization combines a cloud parameterization that infers the presence of clouds based on relative humidity with observations of the cloud mask and cloud top/base height to provide a three-dimensional cloud analysis. During the forecasts, the hydrometeors can be advected and diffused with no microphysics, in what we refer to as the MAD-WRF passive mode. Alternatively, these passive hydrometeors can be integrated into the explicitly resolved hydrometeors during a nudging phase, designated the MAD-WRF active mode (Jiménez et al., 10.1016/j.solener.2022.04.055). As such, MAD-WRF has been extensively used for solar energy predictions.
Here we have investigated the feasibility of using MAD-WRF to improve the accuracy of intense precipitation forecasts. An extreme precipitation event over Israel that led to urban floods and two casualties in Tel-Aviv during January 4th, 2020, has been chosen as a case study. The extreme accumulated precipitation responsible for noon and early afternoon floods was triggered by a persistent cloud train that developed over the area several hours before. MAD-WRF model has been configured with 3-nested domains with 9, 3 and 1 km grid-sizes. We have run MAD-WRF in active mode incorporating satellite-retrieved cloud-top heights provided by the European Space Agency EUMETSAT in all three domains. EUMETSAT data are available in near real-time making it suitable for operational forecasts.
Independent precipitation data measured by the Israel Meteorological Service radar at Bet-Dagan (about 10 km south-east of Tel-Aviv) has been used for forecasts verification. Comparison between radar data and MAD-WRF forecasts with and without incorporation of EUMETSAT cloud-tops retrievals reveal the advantage MAD-WRF cloud initialization. The significant improvement in the forecast of the location and rate of the precipitation is observed up to 12 hours ahead in time.
On-going work focuses on the evaluation of the precipitation distributions and improvement of the forecast of dry areas.
How to cite: Gelman, A., Morin, E., Jiménez, P., Sheu, R.-S., and Rostkier-Edelstein, D.: Improving forecasts of extreme precipitation with MAD-WRF mesoscale model, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-17568, https://doi.org/10.5194/egusphere-egu25-17568, 2025.
EGU25-4905 | ECS | Posters virtual | VPS2
Deep learning-based ENSO modeling and its prediction and predictability studyTue, 29 Apr, 14:00–15:45 (CEST) | vP5.5
A novel deep learning (DL) transformer model, named the 3D-Geoformer, has been developed for ENSO-related modeling studies in the tropical Pacific. Multivariate input predictors and output predictands are selected to adequately represent ocean-atmosphere interactions; so, this purely data-driven model is configured in such a way that key fields for the coupled ocean-atmosphere system are collectively and simultaneously utilized, including three-dimensional (3D) upper-ocean temperature and surface wind stress fields, which represents the coupled ocean-atmosphere interactions known as the Bjerknes feedback in the region. The 3D-Geoformer achieves high correlation skills for ENSO prediction at lead times of up to one and a half years. The reasons for the successful prediction with interpretability are explored comprehensively by performing perturbation experiments to predictors and quantifying input‐output relationships in predictions using the 3D-Geoformer. This is achieved by investigating how the thermal precursors contribute to ENSO prediction skills, with the dependence of the precursor representations on preconditioning multi-month input predictors elucidated. Results reveal the existence of ENSO‐related upper‐ocean temperature anomaly pathways and consistent phase propagations of thermal precursors around the tropical Pacific in the DL framework. The research demonstrates that 3D thermal fields and their basinwide evolution during multi-month time intervals act to enhance long‐lead prediction skills of ENSO. It is demonstrated that the 3D-Geoformer can not only have its ability to effectively improve prediction skills of sea surface temperature variability in the eastern equatorial Pacific, but also explain how and why it is so, thus enhancing model explainability.
How to cite: Zhou, L. and Zhang, R.: Deep learning-based ENSO modeling and its prediction and predictability study, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4905, https://doi.org/10.5194/egusphere-egu25-4905, 2025.
