Accurate classification of precipitation type (convective vs. stratiform) from passive microwave satellite observations is fundamental for understanding global precipitation patterns and improving weather prediction models. While previous studies have demonstrated the effectiveness of deep learning approaches with accuracies above 90% on smaller temporal windows, questions remain about model generalization across extended time periods and different surface types. This study presents a comprehensive analysis using a nine-year Global Precipitation Measurement (GPM) mission dataset, followed by a detailed investigation of surface-type specialization. 

Our primary analysis leverages over 400 million samples from 2014-2022, using 32x32 pixel patches and a ResNet architecture. This large-scale model achieved an accuracy of 85% on a holdout test-year, with balanced performance across both precipitation types (F1-score of 0.85 for both convective and stratiform classes). While matching the general performance range of previous approaches, these results demonstrate robust generalization capabilities across a much longer temporal span and diverse global conditions, using a significantly larger training dataset. 

To further investigate model generalization, we conducted a specialized analysis to examine performance across different surface types creating distinct datasets for land and ocean. ResNet-50 architectures were trained for three comparative models: a baseline model using combined data, a land-only, and an ocean-only model. Analysis revealed that the baseline model achieved robust performance across both surface types (82 % accuracy over land, 86 % over ocean). Surprisingly, surface-specific models showed minimal improvement, with the land-specific model achieving 81% and the ocean-specific model reaching 85% accuracy on their respective domains. This suggests that larger, diverse datasets enable models to learn more robust features that generalize well across data subsets. 

These findings demonstrate that deep learning models can effectively maintain consistent performance when scaling to multi-year global datasets, and suggest that investing in larger, more diverse training datasets provides robust generalization across both temporal and spatial dimensions without requiring specific subsetting. The approach could lead to more efficient operational systems while maintaining reliable classification accuracy across different surface types and extended time periods. 

How to cite: Orescanin, M., Duvio, D., and Petkovic, V.: Large-Scale Deep Learning for Global Precipitation Type Classification, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14109, https://doi.org/10.5194/egusphere-egu25-14109, 2025.

We use machine learning (ML) to map column-integrated water vapor (IWV) to characterize atmospheric rivers (ARs) and to map planetary boundary layer (PBL) height given Global Navigation Satellite Systems (GNSS) radio occultation (RO) data. GNSS RO of the Earth’s atmosphere obtains vertical profiles of microwave refractivity with vertical resolution approaching 100 meters. RO is effectively a water vapor sounder in the lower troposphere and is especially sensitive to vertical gradients associated with the top of the PBL. RO soundings undersample synoptic variability of the atmosphere with severe nonuniformity in spatial and solar angle distribution, making the creation of atmospheric model-agnostic level 3 climatologies a complicated task. ML methods have already shown great promise for mapping RO retrieved quantities in the horizontal and time. We present two applications of ML to RO data, the first to characterize ARs, and the second to map PBL height.

In a mission architecture trade study, we use an ML approach to determine what type of small-satellite constellation would be appropriate to map ARs with detail sufficient for atmospheric process studies and for the prediction of the severe weather on the U.S. Pacific coast that results from ARs. Because ARs are high-volume flows of water vapor in filaments within the PBL and RO sounds water vapor, RO data are ideally suited as input to ML algorithms for the study of ARs. How many low-Earth orbiting RO sounders would be needed to gain desired information on ARs remains an open question, as does our ability to map ARs with existing program-of-record RO data. We answer these questions by formulating ML algorithms to map ARs in the North Pacific Ocean from simulated and real RO data. Simulated RO sounding geolocations are defined by various sizes and types of Walker constellations, realistic GNSS orbits, and the interpolation of refractivity profiles from the ECMWF operational forecast system. We develop two neural networks, one to convert refractivity soundings between 0 and 10 km to IWV and another to map IWV in the horizontal in 1-hr time windows. We find that optimal performance is obtained with Walker constellations of 36 or more RO satellites in near-polar orbits, appropriate orbits for temporal uniformity of the RO’s sampling density. The advent of GNSS RO satellites in micro-satellite and nano-satellite form factors makes such constellations feasible and affordable in the very near future.

We then use ML to map PBL height. The PBL is the part of the atmosphere closest to the Earth’s surface. In the PBL, turbulent processes often affect the vertical redistribution of heat and moisture and their exchange influences cloud evolution and large- to meso-scale circulation. The PBL height can be used to describe climatological processes in a specific region, including cloud characterization. The high vertical resolution of RO observations is suitable to model PBL height in individual profiles. Initially, we use the changes in refractivity to model the PBL height at RO locations and times. Then, we apply ML to produce global PBL heights with a high temporal resolution.

How to cite: Shehaj, E., Leroy, S., and Cahoy, K.: A machine learning framework to map Atmospheric Rivers and Planetary Boundary Layer height from GNSS radio occultation observations, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13623, https://doi.org/10.5194/egusphere-egu25-13623, 2025.

Cloud fraction significantly affects the short- and long-wave radiation. However, its realistic representation in models has been difficult due to inadequate understanding of the sub-grid scale cloud processes. Recently, we have developed a neural network-based scale-adaptive (NSA) cloud-fraction scheme using the CloudSat data and found that the new scheme could greatly improve the simulation of cloud spatial distribution and vertical structure. In this study, we present two applications of the NSA scheme in the WRF model. The first is the simulation of the regional winter climate of the Tibet Plateau, where the NSA scheme was shown to significantly reduce the longstanding bias of too-cold surface temperature. The second is a tropical cyclone simulation, showing that the NSA scheme better simulated the track of In-Fa (2021). The underlying mechanisms will be presented.

How to cite: Chen, G.: The application of a neural network-based scale-adaptive cloud-fraction scheme in the WRF model , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7683, https://doi.org/10.5194/egusphere-egu25-7683, 2025.

A deep ANN-based dimension reduction (DR) method, called DIRESA (distance-regularized Siamese twin autoencoder), has been developed to capture nonlinearities while preserving distance (ordering) and producing statistically independent latent components. The architecture is based on a Siamese twin autoencoder, with three loss functions: reconstruction, covariance, and distance loss. An annealing method is used to automate the otherwise time-consuming process of tuning the different weights of the loss function. DIRESA has been compared with PCA and state-of-the-art DR methods for two conceptual models, Lorenz ’63 and MAOOAM (Modular Arbitrary-Order Ocean-Atmosphere Model), and significantly outperforms them in terms of distance (ordering) preservation KPIs and reconstruction fidelity. The latent components have a physical meaning as the dominant modes of variability in the system. DIRESA correctly identifies the major coupled modes associated with the low-frequency variability of the coupled ocean-atmosphere system. Next to the conceptual model results, the first DIRESA results for reanalysis data will be presented.

DIRESA is provided as an open-source Python package, based on Tensorflow. With one line of code convolutional and/or dense layers DIRESA models can be build. On top of that, the package allows the use of custom encoder and decoder submodels to build a DIRESA model. The DIRESA package acts as a meta-model, which can use submodels with various kinds of layers, such as attention layers, and more complicated designs, such as graph neural networks. Thanks to its extensible design, the DIRESA framework can handle more complex data types, such as three-dimensional, graph, or unstructured data. Its flexibility and robust performance make DIRESA an promising new tool in weather and climate science to distil meaningful low-dimensional representations from the ever-increasing volumes of high-resolution climate data, for applications ranging from analog retrieval to attribution studies.

Tutorial: https://diresa-learn.readthedocs.io/

Preprint: https://arxiv.org/abs/2404.18314

How to cite: De Paepe, G. and De Cruz, L.: DIRESA – A Deep Learning-based, nonlinear "PCA"​, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-6004, https://doi.org/10.5194/egusphere-egu25-6004, 2025.

Weather and climate simulations produce petabytes of high-resolution data that are later analyzed by researchers to understand climate change or severe weather. We propose a new error-bounded compression method targeting for weather and climate data. It contains a JPEG2000 compression layer to capture the bulk part of the data, and a sparse wavelet layer to record the sparse signal that excess the given error bound. The sparse wavelet layer encodes the wavelet coefficients using the SPIHT algorithm. We test our method with established compression methods on a suite of benchmarks including basic statistics, case study of the hurricane data, derived variable computation, and Lagrangian trajectory simulation. Our method is favourable in most benchmark cases at given range relative error targets from 0.1% to 10% achieving compression ratios from 10x to more than 800x. It can reconstruct the derivatives of the compressed data with a best fidelity and does not add high frequency artifacts found in other compression methods. In Lagrangian trajectory simulations, our method can produce less distortion in trajectories and distribution of particles compared with SZ3. We are able to produce a 16x compressed wind data achieving less error metric than adding 5% random noise to the data, making it ready for practical use.

How to cite: Huang, L. and Hoefler, T.: Error bounded compression for weather and climate applications, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-10282, https://doi.org/10.5194/egusphere-egu25-10282, 2025.

Data-driven weather forecasting models, such as FuXi and Pangu-Weather, have made significant advancements in global forecasting accuracy and computational efficiency. However, these models lack physical constraints, a limitation that traditional numerical weather prediction (NWP) models address through the dynamical core and physical parameterization schemes. Recent efforts, like NeuralGCM and PINNs, have successfully integrated the dynamical core or Navier-Stokes equations with machine learning models. Yet, effective integration of physical parameterization schemes remains uncharted in this field, primarily due to the greater uncertainty and complexity of physical processes compared to the dynamical core. To bridge this gap, we integrated the shortwave radiative transfer scheme with FuXi, by modeling the Rapid Radiative Transfer Model for General Circulation Models Applications (RRTMG) as a neural network. This represents the first successful integration of a physical parameterization scheme with large-scale weather forecasting models. This integration yielded substantial improvements in forecasting performance and physical consistency, reducing root mean square error (RMSE) by approximately 15% for radiatively related variables, such as albedo and cloud water mixing ratio, especially for longer lead times.  Moreover, the optimized model demonstrated significantly enhanced atmospheric moisture energy conservation.  This work provides a promising pathway for integrating physical processes into machine learning based weather forecasting models, paving the way for more accurate and physically consistent weather forecasts.

How to cite: huang, Q., zhong, X., fan, X., and li, H.: Physics-Constrained Machine Learning Model Enhances Weather Forecast Accuracy and Physical Consistency, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14355, https://doi.org/10.5194/egusphere-egu25-14355, 2025.

Satellite-based research on global sea surface wind is essential for understanding and monitoring the air-sea interface dynamical processes, driving the need for accurate and efficient assessment. Spaceborne scatterometers, which are among the most relevant sensors for sea surface wind observation, play a vital role in obtaining global ocean surface wind information. However, a significant challenge associated with polar-orbiting satellites lies in data gaps caused by their orbital paths, resulting in missing observations between swaths. While several satellite-derived sea surface wind products have been developed using data assimilation (DA) techniques, these existing methods are time-consuming and require large amounts of diverse data, rendering them computationally expensive. Additionally, the iterative steps of variational algorithms can only perform linear or weakly nonlinear adjustments to the governing equations, which may pose challenges given the highly non-linear nature of these equations.

In this study, we leveraged physics-informed neural networks (PINNs) techniques to reconstruct large-scale sea surface wind fields with sparse scatterometer observations from different satellites, integrating observations and filling gaps. By incorporating physical constraints into the loss function, specifically the Navier-Stokes equations, we efficiently fill the data gaps and reconstruct wind fields that not only match observational data but also adhere to physical principles. Another objective of this work is to introduce the wind speed gradient and direction parallel consistent constraints into the loss function in order to enhance the detail of the reconstructed wind field and increase the accuracy of both wind speed and direction. Structurally, the PINN resembles a fully connected neural network (FCNN), offering the advantage of automatic feature extraction. Our model not only extracts valuable information from existing data but also uncovers complex patterns and correlations in data that are difficult for traditional algorithms to capture. This approach provides a novel perspective and an alternative methodology for wind field reconstruction.

The results show that PINNs can reconstruct wind fields that closely resemble realistic wind patterns, capturing large-scale structures while preserving fine-scale details, thanks to the introduction of wind speed gradient and direction parallel consistent constraints. The training time for this model is about 3 hours, using only a single GPU core. This efficiency is partly due to the fact that PINNs do not rely on ensemble methods or large datasets to produce results. Unlike traditional DA methods, PINN does not depend on an initial best-guess field for assimilating observations. While we use only a small amount of scatterometer observation data, no initial field is required to complete the reconstruction. Additionally, since the PINN represents a continuous and differentiable function, it can produce outputs at any spatial or temporal point within the training domain.

Recognizing their potential for forecast models and data integration, PINNs offer a promising approach for accurate sea surface wind field reconstruction and could serve as an effective alternative to current methods.

How to cite: Bo, R., Zhou, Z., Du, H., Yang, P., Zhao, X., Li, Q., and Zang, Z.: An Alternative Large-scale Sea Surface Wind Field Reconstruction Method Using Sparse Scatterometer Data Based on Physics-informed Neutral Network, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7915, https://doi.org/10.5194/egusphere-egu25-7915, 2025.

Tropical cyclones (TCs) pose signif­icant risks, particularly in coastal regions, making accurate prediction of their track and intensity is crucial for effective disaster preparedness and response. Traditional numerical models struggle with balancing accuracy and computational efficiency, although TC track prediction has achieved substantial progress, challenges remain in forecasting TC intensity, especially rapid intensification (RI). This study aims to (1) develop a Transformer-based deep learning (DL) model to predict TC track, intensity, and 24-hour future intensity change simultaneously, and (2) investigate the relative importance of input variables to the contribution of improving model forecast ability. Based on 2001–2021 best track data and ERA5 reanalysis data over the western North Pacific (WNP), we develop an optimal model called OWZP-Transformer, which leverages the multi-head self-attention mechanism and incorporates the 13 input parameters categorized into four factors: basic, environmental, gradient, and structural. Specifically, our study incorporates structural parameters, which is represented by Okubo-Weiss-Zeta Parameter (OWZP). Our OWZP-Transformer model achieves competitive results for track prediction and shows excellent performance in intensity forecasting for the next 6 hour, with an overall root mean square error (RMSE) of 0.91 m/s. This result represents an improvement of 61.9% to 68.4% compared to existing DL models, which generally have RMSE values above 2 m/s. In addition, our model demonstrates superior performance in predicting 24-hour future intensity change, achieving an overall lower mean absolute error (MAE) of 1.57 m/s, which is 63.5% to 74.6% lower than existing DL models. Furthermore, our model successfully identifies all 11 RI events out of 30 TCs samples from WNP test dataset during 2020-2021. We further evaluate the contributions of each parameter for the first time using two explainable feature importance methods: DeepLIFT and DeepLiftShap. The results indicate that self-contributions and the basic factors play a dominant role in short-term forecasts, while the OWZ parameter plays a significant role following them. This study is the first attempt to comprehensively predict a broad range of TC forecasting tasks using a single DL model, highlighting the potential the OWZP-Transformer model as a reliable tool for enhancing both the accuracy and efficiency of TC predictions.

How to cite: Lin, Z. and Chu, J. E.: Enhancing tropical cyclone track and intensity predictions with the OWZP-Transformer model, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-80, https://doi.org/10.5194/egusphere-egu25-80, 2025.

Understanding and predicting wind flow structures within the atmospheric boundary layer (ABL) across varying stability conditions remains a key challenge in atmospheric science and environmental modeling. Although large-eddy simulation (LES) provides high-fidelity insights, it is computationally prohibitive for near-real-time or large-scale applications. To address this, we propose a deep learning framework that predicts the two-dimensional wind flow fields in the u, v, and w components at a target height z, using corresponding flow fields at three levels above z. By integrating vertical flow correlations and continuity principles, our approach captures essential turbulent features while reducing input dimensionality and eliminating the need for full 3D simulations.

A modified convolutional neural network (CNN) forms the core of this framework, capturing complex spatial and temporal patterns from high-resolution LES datasets. Mass conservation is embedded in the training process to ensure physically consistent results. Preliminary results indicate that the model preserves large-scale turbulence features and captures the influence of higher-elevation dynamics, although smaller high-frequency turbulent features require further refinement. To address this, our ongoing work includes adopting a scale-specific approach to explicitly handle the diverse turbulent length scales observed in the ABL. We are also incorporating multitemporal dynamics and attention mechanisms into the architecture of our model to better account for long-range dependencies over time, thereby enabling the model to adapt to different stability regimes and transitions among them. To enhance interpretability, we will employ explainable AI (XAI) tools such as SHAP and GradCAM, revealing how specific regions in the input influence the emergence of particular turbulent footprints in the predicted flow. These insights guide improvements in both model design and understanding of atmospheric processes governing ABL flow development.

This research underscores the transformative potential of deep learning in boundary layer meteorology. By significantly reducing computational demands while retaining essential flow dynamics, our model enables real-time, high-resolution predictions of ABL flows. This scalable and efficient framework opens new possibilities for diverse applications, including weather forecasting, wind energy optimization, and environmental analysis.

How to cite: Difuntorum, J. K., Katurji, M., Zhang, J., and Zawar-Reza, P.: Multilevel Deep Learning for Non-Local Prediction of ABL Flow Fields Across Varying Stability Regimes, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14604, https://doi.org/10.5194/egusphere-egu25-14604, 2025.

The Arctic Ocean plays a critical role in global climate dynamics, yet direct observations of its surface energy budget are sparse and largely constrained to individual field campaigns. Current estimates often rely on reanalysis data, notably ERA5, which demonstrates systematic biases in boundary-layer properties and surface fluxes over Arctic sea-ice. In this study, we train an artificial neural network (ANN) to predict surface fluxes observed during MOSAiC, SHEBA, Arctic Ocean 2018 and ARTofMELT expeditions using ERA5 data as input. Data from shorter field campaigns, such as N-ICE, are used for testing our model against unseen data. Our results indicate that ERA5 Arctic surface fluxes exhibit very low correlations with observations and are characterized by large RMSE values. Our predictions demonstrate significant error reductions across key variables: sensible heat flux (~39%), 2m temperature (~39%), net shortwave radiation (~37%), downward longwave radiation (~21%) and net longwave radiation (~17%). Furthermore, we find a higher correlation (~0.58) with observations compared to ERA5 (~0.24) and approximately 50% reductions in RMSE of the hourly total surface energy budget, i.e. the sum of the individual fluxes. We produce a bias-corrected estimate of surface energy fluxes over Arctic sea-ice. We use our bias correction to revise previous estimates of the climatological surface energy budget over Arctic sea ice. We will make the trained weights available to allow for the custom derivation of bias-corrected fluxes in individual case studies and for climate model evaluation.

How to cite: Hossain, A., Keil, P., Grover, H., and Pithan, F.: Leveraging observations to bias-correct the Arctic surface energy budget in reanalysis through machine learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-18459, https://doi.org/10.5194/egusphere-egu25-18459, 2025.

Large scale machine learning is currently revolutionizing Earth system modeling and the next generation of models will likely be machine learning-based or contain substantial machine learning components. The lack of a complete equation-based description of the Earth system as well as the availability of a plethora of high-quality data make this a tantalizing possibility to obtain models with unprecedented capabilities. In the first part of the talk, we will discuss grand challenges for building machine learning-based Earth system models and what milestones have already been achieved in the last years. We will also examine cases where machine learning models already surpass state-of-the-art equation-based models, e.g. for medium range weather forecasting. In the second half of the talk, we will introduce the WeatherGenerator project that aims to build a next generation, machine learning-based Earth system model. Led by leading European modeling centers but open-source from day-1, the project will train on a wide range of datasets to build a seamless prediction model that can faithfully represent Earth system dynamics from sub-km scale, short term processes to multi-decadal projects. The project will also consider selected applications to ensure the real-world applicability of the developed model.

How to cite: Lessig, C.: Towards machine learning-based Earth system models, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-17016, https://doi.org/10.5194/egusphere-egu25-17016, 2025.

Fine particulate matter (PM_2.5) and ozone pollution cause a significant number of premature deaths in China each year. Evaluating the effects of emission mitigation and climate change on air quality with climate-chemistry models is computationally expensive. In this study, we develop two machine learning models — a deep learning framework based on U-Net image segmentation and long short-term memory (LSTM) neural networks, and an extreme value model — to predict China’s future air quality trends. These models are trained using model simulation results from 2014 to 2022 under high to low anthropogenic emission scenarios. The deep learning model yields promising results for PM_2.5, with an R² of 0.79 and root mean squared error (RMSE) of 12.58 ug/m³. The extreme value model for ozone episode exceedance rates achieves an R² of 0.97 and RMSE of 1.43%. This work demonstrates the potential of deep learning and extreme value models in efficiently modeling air quality, offering robust tools for future air quality assessments.

How to cite: Wan, F., Shen, L., and Jiang, Z.: Efficient machine learning frameworks for prediciting China’s future air quality trends, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14723, https://doi.org/10.5194/egusphere-egu25-14723, 2025.

Dust aerosol forecasting is of significant scientific and societal importance. Currently, the most accurate forecasting systems rely on numerical weather prediction methods, which solve differential equations to simulate the physical and chemical processes of dust aerosols and predict dust concentrations. However, errors introduced by initial and boundary conditions, along with the complex nonlinear interactions between aerosol physical-chemical processes and atmospheric dynamics, result in uncertainties and high computational costs in numerical prediction methods. In recent years, artificial intelligence (AI) methods have demonstrated significant potential in the field of weather forecasting. However, AI-based approaches for addressing such challenging extreme weather events remain in their infancy. Here, we introduce DustWatcher, an AI-based forecasting method for global dust aerosols. DustWatcher integrates spatiotemporal Transformers with conditional generative networks to develop a neural network framework that optimizes forecasting errors in an end-to-end manner. Compared to current state-of-the-art global and regional aerosol forecasting systems, DustWatcher, trained on 41 years of global reanalysis aerosol data, delivers more accurate deterministic forecasts for most aerosol variables including surface dust concentration and aerosol optical depth. DustWatcher provides skillful forecasts every three hours for the next seven days at a resolution of 0.5°×0.625°. Our results demonstrate the potential of AI in improving the dust forecasts accuracy and advancing its application in dust aerosol forecasting field.

How to cite: Du, S. and Chen, S.: Deep Learning for Accurate Global Dust Aerosol forecasting, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-3633, https://doi.org/10.5194/egusphere-egu25-3633, 2025.

Diffusion models have been shown highly capable for image generation tasks and more recently have been adapted for weather forecasting, allowing sharper predictions than forecasts generated by previous methods, and straight-forward generation of ensemble predictions. Their value for image-based predictions is evident. Dust storms are frequent high-impact weather phenomena in West Africa that directly impact human life, e.g., by disrupting land and air traffic, posing health threats, and affecting energy delivery from solar-energy systems. Timely and precise prediction of these phenomena is crucial to mitigate adverse impacts.

State-of-the-art machine learning-based weather prediction (MLWP) models do not predict dust since they are limited by computational constraints and by the need of high-quality aerosol reanalyses. Moreover, the current operational numerical weather prediction (NWP) models for Africa still need improvement for resolving the short-scale dynamics and surface properties which leads to the formation of convective dust storms, and also often the convection itself. This is where observation-based short term forecasts (“nowcasts”) become particularly valuable. Nowcasts can provide greater skill than NWP on short time-scales, can be frequently updated, and have the potential to predict phenomena currently operational NWP and MLWP models do not reproduce. However, despite routine high frequency and high resolution observations from satellites, as of January 2025, no nowcast of dust storms is available.

In this study, we present an image-based approach for nowcasting dust storms: we apply a diffusion model to predict next frames of the SEVIRI desert dust RGB composite, a product of false-colour satellite images highlighting both dust and deep convection. We create nowcasts of this RGB composite for a large domain over West Africa up to 6 hours ahead and show that our nowcasts can predict both convective storms and convectively generated dust storms which currently operational NWP may not reliably reproduce. Furthermore, we create ensemble predictions, allowing a probabilistic forecast assessment.

Our approach provides a valuable tool that could be used in operational forecasting to improve the prediction of convective storms, dust storms, and indeed other weather events. Due to the technical similarity of RGB composite imagery from geostationary satellites, this approach could also be adapted to nowcast other RGB composites, such as those for ash, or convective storms. In the wider context, such nowcasts of brightness temperatures and brightness temperature differences, which the RGB composites are based on, could be used for predicting other products which use these satellite retrievals.

How to cite: Hermes, K., Marsham, J., Klose, M., Marenco, F., Brooks, M., and Bollasina, M.: Diffusion models for image-based nowcasting of desert dust for West Africa, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13224, https://doi.org/10.5194/egusphere-egu25-13224, 2025.

Machine Learning (ML) has been widely explored for its potential in modelling air quality in numerous studies in the past. However, these approaches approximated function that maps the finite-dimensional input and output vectors. This restricts their extrapolation to unseen data and different discretization. Neural operators, a class of neural networks approximate the operator between infinite dimensional input and output functions. These models learn the underlying operator between input functions and their time-evolved state directly from data. In this work, we introduce our contribution to the field of neural operators termed Complex Neural Operator (CoNO) to learn the evolution of PM2.5 and CO concentrations over India. We trained our models using WRF-Chem simulated data over India for the years 2016-2018 and evaluated it for the year 2019. We assess our models for forecasting high pollution events, long-term forecasting (up to 72 hours) and city-level forecasts for six cities targeting two key pollutants.

How to cite: Bedi, S., Krishnan, N. M. A., and Kota, S. H.: Developing Neural Operators for Modeling PM2.5 and CO over India, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-10865, https://doi.org/10.5194/egusphere-egu25-10865, 2025.

Urban and industrial areas are vulnerable to accidental releases of pollutants. To accurately determine the pollutant's plume position and affected areas, it is essential to estimate the atmospheric flow around the affected site. This flow can be precisely computed using numerical methods of Computational Fluid Dynamics (CFD). However, CFD computation is expensive and slow, making it unsuitable for emergency response. As reduced order approximations, machine learning surrogates offer a promising alternative as they are usually much faster; but they must first be trained on CFD-generated data. In this study, we propose a database of atmospheric simulations with varying meshes and atmospheric stability conditions. Meshes are built by randomly sampling buildings and positioning them in space. For each mesh, values of the Monin-Obukhov length and of the ground roughness are sampled, leading to different turbulent regimes and overall atmoshperic flow behaviour. We then train a MeshGraphNet on this database, i.e. a graph neural network built on the mesh structure. The performance of the trained neural network on unseen scenarios with different initial conditions has been evaluated and will be presented.

How to cite: de Villeroché, A., Le Guen, V., Mouradi, R.-S., Massin, P., Bocquet, M., Farchi, A., Cheng, S., and Armand, P.: MeshGraphNets for 3D atmospheric flow in Urban Environment for Atmospheric Dispersion, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-10297, https://doi.org/10.5194/egusphere-egu25-10297, 2025.

Here we present the joint German-Brazilian project QUALARIA (Artificial Intelligence based system for sub-urban scale air quality prediction/Sistema baseado em Inteligência Artificial  para previsão de qualidade do ar em escala sub-urbana). 

Through joint effort between research and business partners in Brazil and Germany, QUALARIA proposes to develop an operational artificial intelligence-based system for monitoring, simulating and predicting air quality in urban environments with unprecedented spatial resolution availability (https://meteoia.com/qualaria/). New downscaling approaches based on artificial intelligence have recently shown promising performance to simulate sub-grid atmospheric processes. This approach is designed to monitor and predict atmospheric pollutant concentrations and air quality indexes with high spatial resolution, through the development of the QUALARIA system. Advanced global and regional chemical-meteorological models, such as reanalysis data from ERA5 and EAC4, and WRF-Chem simulations are applied to derive the climatological state of air composition, specifically the average levels of air pollutant based on existing emission inventories. Measurements of PM10, PM2.5 NO2, and O3 concentrations from Air Quality Automatic Stations of the Environmental Company of São Paulo State (CETESB) are used to train the downscaling AI algorithm to capture the sub-grid spatial variations of the pollutant concentrations. Low-cost sensors are deployed to increase and complement the spatial coverage of the CETESB network. Artificial intelligence will transform air quality maps at a horizontal resolution of 10 km to street-level maps with an increased resolution of 100 m. From its simulated and predicted downscaled pollutant concentration fields, QUALARIA will provide its users with relevant air quality indicators, informing about the impacts of air pollution in human health and activities via an online dashboard. To achieve the optimal dashboard design, public and private sector stakeholders are being engaged and consulted for the co-development of the dashboard design and features.

How to cite: Li, C. W. Y., Peli, V. M. L., Calderón, M. E. G., Perez, G. M. P., Martin, T. C. M., de Lucena, A. V., Barbosa, E. L. S. Y., Schindler, M., Laimer, F., Andrade, M. D. F., Dias de Freitas, E., and Brasseur, G.: An AI System to Predict Street-Level Air Quality: Introduction to the QUALARIA Project, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4569, https://doi.org/10.5194/egusphere-egu25-4569, 2025.

The applications of deep learning (DL) technique in atmospheric environment research are expanding rapidly. Here we developed a DL framework to quantify the responses of surface ozone (O₃) to anthropogenic and meteorological changes in China. The DL-based analysis suggests volatile organic compound (VOC)-limited regimes in urban areas over northern inland China, and thus, reductions of nitrogen oxide (NO_x) emissions have resulted in increases in surface O₃ concentrations. In contrast, changes in meteorological conditions led to a dramatic decrease in surface O₃ concentrations in 2019-2021, particularly, in the North China Plain, whereas the decline in surface O₃ concentrations driven by beneficial meteorological conditions in 2019-2021 has been completely reversed due to the occurrence of long-lasting heatwave in 2022, particularly in central China. The DL framework, developed in this work, provides a novel data-driven pathway to assess the causes of surface O₃ changes, and is helpful for a comprehensive understanding of the driving factors of surface O₃ evolution in China.

How to cite: Jiang, Z., Chen, X., Wang, M., and He, T.-L.: Deep learning-based surface O3 responses to anthropogenic and meteorological changes, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-2638, https://doi.org/10.5194/egusphere-egu25-2638, 2025.