For understanding localized hydrological and climatological processes, downscaling gridded precipitation data to finer spatial resolutions is a crucial prerequisite. For a densely populated country like India, accurate downscaled data is crucial for building resilience to climate change impacts, supporting adaptation efforts, and enhancing disaster management. In recent years, deep learning (DL) has emerged as a powerful tool for advancing Earth system modelling and climate data downscaling. This study presents a comprehensive intercomparison of deep learning architectures, for downscaling precipitation across India. A few efficient DL architectures from recent studies are chosen for intercomparison such as simple dense, simple convolutional neural network, Fast Super Resolution Convolutional Neural Network (FSRCNN), Super Resolution Deep Residual Network (SRDRN), U-Net, and Nest-U-Net. The experiments are designed in synthetic style by using coarsened ECMWF Reanalysis version 5 (ERA5; 1ox1o) daily variables as the inputs and high-resolution Indian Monsoon Data Assimilation and Analysis reanalysis (IMDAA; 0.12ox0.12o) daily precipitation as training labels and benchmarks for the evaluation. Training and validation are conducted for the period 1980-2014, afterwards the trained models are evaluated on data from 2015-2020. To reduce the biases induced by the highly positive-skewed precipitation data and to enhance the model performance on extreme events, a weighted mean absolute error is implemented for training. The performance of the DL models is also compared with the Bias Correction and Spatial Disaggregation (BCSD), a renowned statistical downscaling method. The results indicate that all deep learning DL models outperformed the BCSD method. Among the DL models, U-Net and Nest-U-Net demonstrated superior performance in capturing fine-scale precipitation patterns and extreme precipitation events, owing to their encoder-decoder architecture, which effectively learns spatial features at different scales. In contrast, the FSRCNN and SRDRN produced results with slightly lower precision than the U-Net models, but at a significantly reduced inference time, making them more efficient for faster data generation. The findings underscore the potential of deep learning for improving regional precipitation downscaling across India, offering a promising alternative to traditional statistical methods like BCSD in handling complex, non-linear relationships inherent in climate data.
How to cite: Murukesh, M. and Kumar, P.: Comparative analysis of deep learning architectures trained for downscaling gridded precipitation across India, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-676, https://doi.org/10.5194/egusphere-egu25-676, 2025.
Climate change is profoundly affecting ecosystems and societies. Impacts on hydrological regimes, water resources, and urban heatwaves are particularly important, emphasizing the need for a detailed understanding of these changes at local scales to inform effective adaptation strategies. Achieving this requires reliable, high-resolution projections of future climate conditions. However, current climate models operate at coarse spatial resolutions, limiting their ability to capture small-scale processes and extreme weather events. To bridge this gap, robust downscaling techniques are essential for refining the outputs of global and regional climate models.
We propose a multivariate super-resolution (SR) approach to downscale temperature and precipitation data in Switzerland to improve the representation of localized patterns, particularly in Alpine regions, while simultaneously capturing the interdependencies between temperature and precipitation, which are crucial for hydrological applications. We leverage advanced machine learning techniques, including Generative Adversarial Networks (GANs) and Diffusion models, to overcome the limitations of classical methods in capturing inter-variable dependencies. These models provide an ensemble framework, providing multiple possible realizations, to account for downscaling uncertainties, resulting in more robust and reliable outputs for impact modeling and decision-making. We test different loss functions, like a regional CRPS, to allow for variability in the generated meteorological fields.
We compare the performance of GANs and Diffusion models along with the differences between univariate and multivariate settings. Our approach includes applying a multivariate bias correction prior to downscaling. The downscaled results are compared to a setting based on univariate bias correction. Additionally, we present the pipeline, which integrates bias correction and downscaling and is intended to be open source.
How to cite: Horton, P., Samarin, M., Otero, N., Allen, S., and Volpi, M.: Multivariate climate downscaling using deep learning models, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15363, https://doi.org/10.5194/egusphere-egu25-15363, 2025.
Databases of high-resolution interpolated climate data are essential for analyzing the impacts of past climate events and for developing climate change adaptation strategies for managed and natural ecosystems. To enable such efforts, we contribute an accessible, comprehensive database of interpolated climate data for Africa that includes monthly, annual, decadal, and 30-year normal climate data for the last 120 years (1901 to present) as well as multi-model CMIP6 climate change projections for the 21st century. The database includes variables relevant for ecological research and infrastructure planning, and comprises more than 25,000 climate grids that can be queried with a provided ClimateAF software package. In addition, 30 arcsecond (~1km) resolution gridded data, generated by the software, are available for download (https://tinyurl.com/ClimateAF). The climate grids were developed with a three-step approach, using thin-plate spline interpolations of weather station data as a first approximation, subsequent fine-tuning with deep neural networks to capture medium-scale local weather patterns, and lastly dynamic lapse-rate based downscaling to a user-selected resolution, or to scale-free point estimates with the ClimateAF software package. The study contributes a novel deep learning approach to model orographic precipitation, rain shadows, lake and coastal effects, including the influences of wind direction and strength. The climate estimates were optimized and cross-validated with a checkerboard approach to ensure that training data was spatially distanced from validation data. We conclude with a discussion of applications and limitations of this database.
How to cite: Namiiro, S., Hamann, A., Wang, T., Castellanos-Acuña, D., and Mahoney, C.: Climate data interpolation with deep neural networks: a comprehensive dataset of historical and future climate for Africa, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4595, https://doi.org/10.5194/egusphere-egu25-4595, 2025.
Assessing precipitation impacts due to anthropogenic climate change relies on accurate and high-resolution numerical Earth system model (ESM) simulations. However, such simulations are computationally too expensive, and their discretized formulation can introduce systematic errors. These can, for example, lead to an underestimation of spatial intermittency and extreme events.
Generative machine learning has been shown to skillfully downscale and correct precipitation fields from numerical simulations [1].
However, these approaches require separate training for each Earth system model, making corrections of large ESM ensembles computationally costly.
Here, we follow a diffusion-based approach [2] by training an unconditional generative consistency model [3] on high-resolution ERA5 precipitation data. Once trained, a single generative model can be used to efficiently downscale arbitrary ESM simulations in an uncertainty-aware and scale-adaptive manner. Using three different climate models, GFDL-ESM4 [4], POEM [5], and SpeedyWeather [6], we evaluate the performance and generalizability of our approach.
[1] Harris, L., McRae, A.T., Chantry, M., Dueben, P.D. and Palmer, T.N., 2022. A generative deep learning approach to stochastic downscaling of precipitation forecasts. Journal of Advances in Modeling Earth Systems, 14(10), e2022MS003120.
[2] Hess, P., Aich, M., Pan, B., and Boers, N., 2024. Fast, Scale-Adaptive, and Uncertainty-Aware Downscaling of Earth System Model Fields with Generative Machine Learning. arXiv preprint arXiv:2403.02774.
[3] Song, Y., Dhariwal, P., Chen, M., and Sutskever, I. 2023. Consistency Models. In International Conference on Machine Learning (pp. 32211-32252).
[4] Dunne, J.P., Horowitz, L.W., Adcroft, A.J., Ginoux, P., Held, I.M., John, J.G., Krasting, J.P., Malyshev, S., Naik, V., Paulot, F. and Shevliakova, E., 2020. The GFDL Earth System Model version 4.1 (GFDL‐ESM 4.1): Overall coupled model description and simulation characteristics. Journal of Advances in Modeling Earth Systems, 12(11), e2019MS002015.
[5] Drüke, M., von Bloh, W., Petri, S., Sakschewski, B., Schaphoff, S., Forkel, M., Huiskamp, W., Feulner, G. and Thonicke, K., 2021. CM2Mc-LPJmL v1.0: biophysical coupling of a process-based dynamic vegetation model with managed land to a general circulation model. Geoscientific Model Development 14, 4117–4141.
[6] Klöwer, M., Gelbrecht, M., Hotta, D., Willmert, J., Silvestri, S., Wagner, G.L., White, A., Hatfield, S., Kimpson, T., Constantinou, N.C. and Hill, C., 2024. SpeedyWeather.jl: Reinventing atmospheric general circulation models towards interactivity and extensibility. Journal of Open Source Software, 9(98), 6323.
How to cite: Hess, P., Aich, M., Pan, B., and Boers, N.: Downscaling precipitation simulations from Earth system models with generative machine learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-9742, https://doi.org/10.5194/egusphere-egu25-9742, 2025.
Reducing climate model biases is crucial for decreasing uncertainties in future climate projections. Despite recent efforts, improvements between the latest generations of Earth System Models (ESMs) have been modest, primarily due to the continued reliance on subgrid-scale parametrizations. These parametrizations are necessary because the model resolutions in CMIP6 are too coarse to explicitly simulate too small-scale processes such as ocean mesoscale eddies and deep atmospheric convection, which significantly influence regional and global climate patterns. Recent advances in computational power have enabled higher-resolution models, allowing for some of these processes to be simulated explicitly, reducing the need for parametrization. Here, we combine a convolutional neural network (CNN) classifier and explainable AI (XAI) to investigate the role of increased resolution in simulating winter surface temperature fields. The CNN is used to classify ESMs with varying resolutions based on snapshots of their surface temperature fields, while the XAI approach explains which regions and features the CNN relies on to make these distinctions, providing deeper insights into ESM performance. Results indicate that models with similar ocean grids are more frequently confused by the CNN than those from similar modeling centers, emphasizing the crucial role of ocean resolution, particularly the presence of mesoscale eddies, in shaping climate simulations. Although the analysis is restricted to surface air temperature, the XAI approach offers a more nuanced understanding of model differences compared to traditional bias analyses. This methodology can be extended to other climate variables and ESM features, offering a powerful tool for enhancing model intercomparison and evaluating ESM performance.
How to cite: Michel, S., Strommen, K., and Christensen, H.: Unravelling the role of increased model resolution on surface temperature fields using explainable AI, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15256, https://doi.org/10.5194/egusphere-egu25-15256, 2025.
The use of machine learning (ML) for climate science has attracted considerable attention within the last few years. A number of recent studies have used ML to extract information from global climate data (e.g. regional downscaling), predict future states of the climate system and evaluate models against observations. In particular, Brunner and Sippel (2023) showed that low-resolution global climate models and observations can reliably be distinguished based on the global distribution of daily temperature, even after removing the mean model bias. ML is thus able to isolate fundamental differences between models and observations even in the presence of substantial internal variability. This raises the questions of whether ML can also distinguish between model and observational data on a regional scale, whether ML is as successful for km-scale models as for coarse-resolution models, and whether more complex bias correction methods reduce the success of ML.
To answer these questions, we use daily temperature fields over Austria, a topographically very complex domain. As training data, we use 200 different, randomly drawn days from each of the 13 ÖKS15 bias-corrected EURO-CORDEX models with an output resolution of 1km, resulting in 2600 samples labeled “model” which are matched by the same number of random days labeled “observation” from the SPARTACUS observation dataset. We use the binary classification approach to distinguish between the two classes of models versus observations. A logistic regression classifier is trained to determine the probability that a daily temperature field belongs to one of the two classes. In order to evaluate the ML algorithm subsequently, all days from the out-of-sample 10-year period 2005-2014 are used as test data.
The ML algorithm succeeds in correctly identifying the overwhelming majority of the test data for the setup used, resulting in an accuracy of 99%. The results remain consistent even when a different sample of 2x2600 random training days is used. In contrast to more complex classifiers, such as a convolutional neural network (CNN), the learned coefficients from the logistic regression allow insights into the spatial patterns that are crucial for distinguishing between models and observations. While the performance of climate models is typically evaluated on climatological timescales, our results highlight that such classifiers can be used to identify patterns of structural model biases. Our method hence offers a computationally efficient approach for model evaluation, especially when handling km-scale climate model data on a regional domain.
References:
Brunner L. and Sippel S. (2023): Identifying climate models based on their daily output using machine learning, Environmental Data Science, https://doi.org/10.1017/eds.2023.23
How to cite: Meindl, M., Voigt, A., and Brunner, L.: Using machine learning to distinguish km-scale climate models and observations on a regional scale, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-6462, https://doi.org/10.5194/egusphere-egu25-6462, 2025.
Aerosols affect the Earth’s energy budget by both scattering and absorbing solar radiation. Measuring parameters that separately quantify the two components, such as the aerosol absorption optical depth (AAOD), is key to better understanding the aerosol direct climate effect. As most satellite instruments can only retrieve the total aerosol extinction signal, the most reliable source of global AAOD observations is the ground-based AERONET sensor network. AERONET comprises hundreds of stations worldwide; however, their spatial distribution is uneven and coverage remains sparse in many relevant regions. To effectively reduce our uncertainty related to absorbing aerosols and efficiently expand the network, new stations should be placed in locations that maximise measurement informativeness. In this study, we address the problem of optimal sensor placement using convolutional neural processes (ConvNPs). ConvNPs are meta-learning models that use convolutional neural networks to learn maps from heterogeneous input datasets to a context-dependent Gaussian predictive model. We train ConvNPs using reanalysis data to learn to model daily global AAOD from sparse point observations given at station locations and additional gridded auxiliary data. The model’s probabilistic predictions are then harnessed in an active learning framework to sequentially propose new observation locations that optimally reduce model uncertainty and improve the network's informativeness. Our subsequent analysis considers further practical factors that might trade off with informativeness in the selection of new station locations, such as cloudiness and remoteness. The resulting proposed placements identify locations that would optimally enhance ground-based AAOD observation and can inform and focus future network expansion efforts.
How to cite: Pelucchi, P., Coca-Castro, A., Andersson, T. R., Vicent Servera, J., and Camps-Valls, G.: Optimal Sensor Placement for Aerosol Absorption Optical Depth with Convolutional Neural Processes, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-18735, https://doi.org/10.5194/egusphere-egu25-18735, 2025.
In recent years, machine learning has emerged as a promising alternative to numerical weather prediction models, offering the potential for cost-effective and accurate forecasts. However, a significant limitation of current machine learning methods for weather forecasting is the lack of principled and efficient uncertainty quantification—a key element given the complexity of the Earth's climate system and the challenges in modeling its processes and feedback mechanisms. Inadequate uncertainty quantification and reporting undermines trust in and the practical use of current weather forecasting methods (Eyring et al., 2024).
Uncertainty quantification methods for weather forecasting typically use prediction intervals and can be categorized into Bayesian and frequentist approaches. Bayesian methods, while theoretically appealing, often involve restrictive assumptions and do not scale well to the complexity of spatio-temporal data. Frequentist approaches, such as ensemble-based methods, are widely used in weather forecasting and include techniques like perturbing initial states with noise (Bi et al., 2023; Scher et al., 2021), varying neural network parameters (Graubner et al., 2022), or training generative models (Price et al., 2023). However, most frequentist methods provide only asymptotically valid prediction intervals, which may not suffice in all weather forecasting applications.
Conformal prediction (CP) is a promising uncertainty quantification framework that delivers valid and efficient prediction intervals for any learning algorithm, without requiring assumptions about the underlying data distribution (Vovk et al., 2005). Despite its growing popularity in the machine learning and statistics communities, traditional CP methods are not tailored to spatio-temporal data in weather forecasting. This is due to challenges arising from spatial and temporal dependencies—such as spatial autocorrelation and temporal dynamics—that violate the exchangeability assumption underlying standard CP methods. Several recent studies attempted to address these challenges by introducing new CP algorithms specifically designed for various types of non-exchangeability (Oliveira et al., 2024). However, these adaptations face several limitations, including high computational complexity, asymptotic guarantees, and/or the need for recalibration of prediction intervals.
In this presentation, we will evaluate CP methods in the context of weather forecasting and discuss several limitations. In addition, we will highlight recent advances and discuss potential future directions that could address challenges underlying the use of CP in weather forecasting.
References:
Eyring, V., et al. Pushing the Frontiers in Climate Modelling and Analysis with Machine Learning. Nature Climate Change, 2024.
Leutbecher, M., et al. Ensemble Forecasting. JCP, 2008.
Bi, K., et al. Accurate Medium-range Global Weather Forecasting with 3D Neural Networks. Nature, 2023.
Scher, S., et al. Ensemble Methods for Neural Network-based Weather Forecasts. JAMES, 2021.
Graubner, A., et al. Calibration of Large Neural Weather Models. NeurIPS, 2022.
Price, I., et al. Probabilistic Weather Forecasting with Machine Learning. Nature, 2025.
Vovk, V., et al. Algorithmic learning in a random world. New York: Springer, 2005.
Oliveira, R.I., et al. Split Conformal Prediction and Non-exchangeable Data. JMLR, 2024.
How to cite: Mortier, T., Decancq, C., Sale, Y., Javanmardi, A., Waegeman, W., Hüllermeier, E., and Miralles, D. G.: Valid Prediction Intervals for Weather Forecasting with Conformal Prediction, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-17783, https://doi.org/10.5194/egusphere-egu25-17783, 2025.
Subseasonal-to-seasonal (S2S) forecasts are crucial for decision-making and early warning systems in extreme weather. However, the chaotic nature of atmospheric dynamics limits the predictive skill of climate models on S2S timescales. Teleconnections can provide windows of improved predictability, but leveraging these external drivers to enhance S2S forecast skill remains challenging. This study introduces a spatio-temporal neural network (STNN) designed to predict weekly North Atlantic European (NAE) weather regimes at lead times of one to six weeks during boreal winter. The STNN integrates a stacked vision transformer (ViT) encoder and a long short-term memory (LSTM) decoder to capture short- and medium-range variability. By incorporating spatio-temporal data on the stratospheric polar vortex, tropical outgoing longwave radiation, and 1D NAE regime time series, the network can access patterns linked to teleconnections of key drivers of European winter weather. Its modular design enables the application of mechanistic interpretability, providing novel neuron-level insights into the prediction behavior. The improved predictive skill beyond lead week three and enhanced accuracy for specific regimes suggest novel learned patterns of external drivers. Using Activation Maximization (AM), we analyze these learned representations, and by incorporating gradient-based explanations of correct predictions, we infer additional insights into prevalent teleconnections.
How to cite: Bommer, P. L., Kretschmer, M., Spurler, F., Bykov, K., Boehnke, P., and Hoehne, M. M.-C.: Combining spatio-temporal neural networks with mechanistic interpretability to investigate teleconnections in S2S forecasts, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-9753, https://doi.org/10.5194/egusphere-egu25-9753, 2025.
Clouds influence Earth’s climate by reflecting sunlight and trapping heat, but their role in climate change remains uncertain, causing major unpredictability in models. Global 3D cloud data can improve predictions.
Observations from NASA’s CloudSat mission have advanced our understanding of cloud structures but are limited by long revisit times and narrow coverage. Imaging instruments offer broader, faster coverage but lack vertical information.
In [1] a deep learning approach addressed this challenge by combining MSG/SEVIRI satellite imagery with CloudSat profiles to extrapolate vertical cloud structures beyond observed tracks. Using geospatially-aware Masked Autoencoders, models were pre-trained on a year of MSG data (2010) and fine-tuned with CloudSat tracks as ground truth. This self-supervised training improved reconstruction, outperforming previous methods and simpler architectures [2].
In this work, we explore to what degree including information of the temporal dynamics of clouds can further improve the quality of the 3D cloud reconstruction. Instead of using a single image as input we use a temporal sequence of MSG/SEVIRI images, spanning a period of several hours before and after the target cloud vertical profile. We use a combination of the geospatial encodings used in [1] and the temporal encoding used in [3] to embed these spatiotemporal MSG/SEVIRI cubes in rich, general purpose latent space. We then use a finetuning model as in [1] to map the embeddings into 3D radar reflectivity maps.
We perform a sensitivity analysis to explore how the quality of the reconstruction varies as a function of the amount of temporal information included. We also explore the relative strengths of different pre-training strategies with respect to the quality of the 3D reflectivity reconstruction and cloud type segmentations. With this, we provide insights on self-supervised learning for atmospheric applications.
References
- Stella Girtsou et al. “3D Cloud reconstruction through geospatially-aware Masked Autoencoders” 2024. arXiv: 2501.02035 [cs.CV]. URL: https://arxiv.org/abs/2501.02035.
- Sarah Brüning et al. “Artificial intelligence (AI)-derived 3D cloud tomography from geostationary 2D satellite data”. en. In: Atmos. Meas. Tech. 17.3 (Feb. 2024), pp. 961–978.
- Yezhen Cong et al. SatMAE: Pre-training Transformers for Temporal and Multi-Spectral Satellite Imagery. 2023. arXiv: 2207.08051 [cs.CV]. URL: https://arxiv.org/abs/2207.08051.
How to cite: Diaz, E., Bintsi, K.-M., Castiglioni, G., Eisinger, M., Freischem, L., Girtsou, S., Johnson, E., Jones, W., Jungbluth, A., and Massant, J.: Reconstructing 3D vertical cloud profiles using cloud dynamics, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13734, https://doi.org/10.5194/egusphere-egu25-13734, 2025.
Machine learning (ML) parameterisations for climate models are emerging as a promising approach for capturing subgrid-scale processes, which are not explicitly resolved in climate models due to limitations on resolution. These ML parameterisations are typically trained on datasets generated by high resolution climate models or existing parameterisations (“offline”), but evaluated based on their performance when coupled into an existing climate model (“online”). Quantifying uncertainties associated with ML parameterisations is crucial for gaining insights into the reliability of hybrid ML-climate models.
I will discuss uncertainties associated with an ML parameterisation for atmospheric GWs, focusing on the parametric uncertainties which originate during the training process. I will show how these can propagate when coupled online, becoming a significant source of uncertainty in climate model circulation that we must consider carefully when building ML parameterisations.
How to cite: Mansfield, L. and Sheshadri, A.: Uncertainty Quantification of Machine Learning Parameterisations, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-12129, https://doi.org/10.5194/egusphere-egu25-12129, 2025.
Machine learning (ML), and deep learning (DL) in particular, hold the potential to solve long-standing challenges in understanding and modeling the Earth system. Earth system model (ESM) development is reluctant to implement DL algorithms because they are considered intransparent, meaning it is unclear how these models extrapolate to unseen conditions, e.g., under a changing climate. Still, machine learning is often used to extrapolate into the future, which can lead to misleading results.
We demonstrate these limitations and the dangers of performing naive extrapolation by using a set of deep neural networks to emulate simulated data of gross primary production (GPP). We use a process-based model (PBM) that simulates photosynthetic CO2 uptake as a product of radiation (PAR), stress from daily meteorology (fTmin , fVPD , fSM ), vegetation state (fPAR), and CO2 (εmax(CO2 )). It is given by
GPP = εmax (CO2 ) · PAR · fPAR · fTmin · fVPD · fSM + ε. (1)
The PBM contains many of the typical challenges when using ML for Earth’s system science. It accounts for stochastic noise (ε), is capable of exhibiting multi-year memory, and the predictors are highly correlated on multiple time scales. Further, this model exhibits interesting extrapolation behavior as some of the factors (fTmin , fVPD , fSM , fPAR) saturate in extreme meteorological conditions while others (PAR, εmax ) do not. We feed the PBM with predictors obtained from historical and future climate simulations of a comprehensive Earth system model. The training dataset contains the predictors and predictions of the PBM for various locations in a similar climate zone but different continents and for the historical time frame (1850-present) together with a spurious predictor, namely, surface wind speed. To obtain a set of independent models, each of the co-authors separately implements a custom architecture, without knowing which predictor is which. This results in four different models, namely a linear model, a multi-layer perceptron, a long-short term memory (LSTM), and an attention-based model.
We find that all models show strong prediction performance in cross-validation (Normalized Nash–Sutcliffe Efficiency (NNSE) > 0.9), decent performance when extrapolating to sites on different continents (NNSE > 0.7), but three out of four models show virtually no skill when predicting to a changed climate (NNSE < 0.6). Additionally, most models emit gradients in the same order of magnitude as the PBM when ignoring values where some factors saturate. This indicates that the networks did not learn the saturation behavior from the data. Further, the model that extrapolates best is the LSTM, a model that has a built-in maximum output and, hence, has to saturate.
In conclusion, strong spatial generalization and cross-validation performance do not guarantee decent extrapolation for neural networks even in relatively simple, stable systems. These findings highlight the importance of selecting architectures in line with the expected extrapolation behavior when predicting Earth’s system processes under climate change conditions.
How to cite: Reimers, C., ElGhawi, R., Kraft, B., and Winkler, A. J.: Limitations of Machine Learning Models in Extrapolating to a Changing Climate, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-16360, https://doi.org/10.5194/egusphere-egu25-16360, 2025.
Making projections of possible future climates with models is essential to improve our understanding of the causes and implications of anthropogenic climate change. While Earth system models are currently the most complete description of the Earth system, these models are computationally expensive. Simpler models (emulators) are therefore useful to explore the large space of possible future climate scenarios and to generate large ensembles. One class of emulators are simple climate models (SCMs), which model the Earth system with simplified physics. A second class of emulators are statistical models, which learn relationships directly from correlations in climate model data. In this preliminary work, we seek to combine the benefits of the physical grounding of SCMs with those of purely statistical emulators, using tools from causal representation learning. The resulting causal climate emulator may allow exploration of the effect of various interventions on the Earth system, including the effect of changing forcings.
The goal of causal representation learning (CRL) is to simultaneously learn low-dimensional latent representations from high-dimensional data, and a causal graph between these latent representations. In the context of climate model data, we aim to infer latent variables representing regions with shared climate variability from fine-grid climate model data, and causal teleconnections between these regions, representing climate dynamics. We build on recent previous work by Boussard et al., which illustrated how a CRL method, Causal Discovery with Single-parent Decoding (CDSD), may be used for this task. CDSD is a continuous optimization method to learn a distribution over latent variables such that every grid-point observation is driven by a single latent variable, and a causal graph between these latents is also learned.
We illustrate that on surface fields of monthly pre-industrial climate model data, CDSD learns physically-reasonable latent variables but learning a robust causal graph between the latent variables remains a challenge. We evaluate our models on synthetic data that approximate the spatiotemporal structures that we observe in climate model data. By autoregressively rolling out the model we can then generate an ensemble of future climate trajectories with the learned generative model. We develop a Bayesian filter to maintain a constant spatial spectrum throughout our autoregressive rollout, and show that it leads to stable climate prediction. Finally, we explore approaches for including the effect of forcings such as greenhouse gasses in the model.
How to cite: Boussard, J., Hickman, S., Trajkovic, I., Kaltenborn, J., Gurwicz, Y., Nowack, P., and Rolnick, D.: Causal climate emulation, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13307, https://doi.org/10.5194/egusphere-egu25-13307, 2025.
Neural operators are transforming computationally intensive scientific disciplines such as weather forecasting and climate modeling, accelerating simulations by several orders of magnitude. However, they often fail to respect fundamental physical principles, such as conservation laws, during long autoregressive rollouts. We introduce an efficient correction layer that enforces global conservation constraints in neural operators. For initial conditions approximately satisfying the constraints, we prove that conservation can be guaranteed while only moderately increasing the total runtime. In a number of fluid dynamics experiments, our method produces physically realistic simulations while maintaining the computational advantages of neural operators. Our results enable the development of reliable and efficient climate model emulators by ensuring that crucial physical balance equations, such as mass and energy, are preserved during extended simulations.
How to cite: White, A., Duruisseaux, V., Bonev, B., Azizzadenesheli, K., Anandkumar, A., and Boers, N.: Enforcing Conservation Laws in Neural Operators for Earth System Modeling, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-19180, https://doi.org/10.5194/egusphere-egu25-19180, 2025.
We consider a data-driven framework for prediction tasks in which the dynamics are learnt in a low-dimensional latent space. We rely on dimensionality reduction techniques --- linear principal component analysis and nonlinear autoencoders and their variants --- to then learn dynamical evolution in the corresponding latent space using disparate methodologies --- linear inverse modeling, dictionary-based sparse regression, reservoir computing, neural differential equations, attention-based transformers, etc. In this setting, we seek to better understand the interplay between the spatial and temporal representations of variability and how they affect prediction skill.
Balu Nadiga, Los Alamos National Laboratory and Kaushik Srinivasan, University of California Los Angeles
How to cite: Nadiga, B. and Srinivasan, K.: Latent Space Prediction of Dynamical Systems, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-21270, https://doi.org/10.5194/egusphere-egu25-21270, 2025.
North Atlantic sea surface temperatures (NASST), particularly in the subpolar region, exhibit some of the highest predictability across global oceanic systems. However, the relative contributions of atmospheric versus oceanic influences on the long term NASST variability remains ambiguous. In this study, we utilize neural networks (NNs) to assess the significance of various atmospheric and oceanic predictors in forecasting the state of NASST within the CANARI Large Ensemble, which employs the Met Office CMIP6 physical climate model (HadGEM3-GC3.1) at a high-resolution atmospheric scale (N216, approximately 60 km at midlatitudes) and a 1/4° resolution for oceanic data. The ensemble comprises forty members, driven by CMIP6 historical data and SSP3-7.0 scenarios for the period from 1950 to 2099. First, we evaluate the ability of the NNs to anticipate the phases of long term (multidecadal variability) using observational datasets, thereby investigating the consistency of physical processes influencing NASST variability between modeled predictions and real-world observations. Second, the research delves into how the interplay between oceanic and atmospheric predictors, alongside external forcings and internal variability (atmospheric noise), impacts the machine learning-based predictions and we use explainable AI techniques to identify the sources of predictability and to pinpoint physical mechanisms and regions crucial for accurate NN forecasts.
How to cite: Colfescu, I.: Explainable neural nets for disentangling sources of predictability in the North Atlantic Sea Surface Temperature (NASST), EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-6680, https://doi.org/10.5194/egusphere-egu25-6680, 2025.
Earth system models (ESMs) are widely used to study climate changes resulting from both anthropogenic and natural perturbations. Over the past years, significant advances have been made through the development of new numerical schemes, refined physical parameterizations, and the use of increasingly powerful computers. Despite these advances, tuning ESMs to accurately reproduce historical data remains largely a manual process, and persistent errors and biases continue to challenge their accuracy. Reducing uncertainties in long-term climate projections and accurately estimating the spread of climate simulations continue to be critical challenges.
Recent advances in machine learning have motivated the development of learning-based methods for the calibration of ESMs. One emerging area of research is the design of hybrid modeling approaches, which combine a physical core with a machine learning model. Training these hybrid models end-to-end (or online) has the potential to unify various challenges in ESMs development, ranging from building subgrid scale parameterizations, to bias correction and parameter tuning.
Training hybrid models online requires working with an optimization problem that depends on the numerical integration of the system. Solving this optimization problem using gradient-based approaches requires the system to be differentiable, or to have access to the adjoint of the numerical model, which is not the case for most of the large-scale physical models. Beyond the need for differentiability, developing hybrid models requires interfacing a physical core that is implemented in low-abstraction languages that are running on CPUs, with AI-based models that are developed using high-abstraction, rapidly evolving languages that run on GPUs. While this interface is not a problem at inference time, doing this interface at calibration time, which is necessary when doing online learning, is not trivial as it would require an iterative communication between components that are implemented on different architectures.
In this work, we aim to investigate online learning and hybrid models to develop new computing paradigms, tools, and calibration methods for designing numerical models that are closely aligned with observations. We study the potential of online learning for deriving efficient and scalable solutions to the above-mentioned problems for applications that include both short-term forecasting and long-term simulations, which require stability considerations of the resulting hybrid systems. We explore learning configurations that include both fully differentiable and black-box physical cores. The latter configuration aims at evaluating the extent to which differentiable programming frameworks can upscale modeling capabilities in terms of accuracy, computational efficiency, and adaptability to represent diverse physical processes.
How to cite: Ouala, S., Meunier, E., Fablet, R., and Le Sommer, J.: Leveraging Differentiable Programming and Online Learning for the design of Hybrid Numerical Models, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13717, https://doi.org/10.5194/egusphere-egu25-13717, 2025.
Atmospheric ozone is a crucial absorber of solar radiation and an important greenhouse gas. However, explicitly representing ozone in climate models is computationally expensive. A recent study introduced a simple linear machine learning-based ozone parameterization scheme (mloz) for daily ozone prediction based on temperature. Here we develop and implement the mloz in the UK Earth System Model (UKESM) for long-term idealized climate simulations. It produces stable ozone predictions over 50 years with a computational cost of less than 0.5% of the total runtime. The scheme accurately predicts ozone distribution, with climatology field errors of less than 10% in the stratosphere. It also realistically represents ozone variabilities, including seasonal and Quasi-Biennial Oscillation-related variabilities, despite a slight underestimation of amplitudes over the stratospheric polar regions. Additionally, we further demonstrated its generalizability by successfully transferring the mloz trained on UKESM to the ICOsahedral Nonhydrostatic model (ICON). Over 30 years of climate sensitivity tests indicate that it can effectively represent the response of ozone to the sudden quadrupling of CO2, significantly outperforming the simplified linearized ozone photochemistry scheme (Linoz) in the troposphere. This implies that the mloz can be transferred to other climate models without a full chemistry module to enable an efficient explicit ozone simulation.
How to cite: Ma, Y., Abraham, L., Versick, S., Ruhnke, R., Braesicke, P., and Nowack, P.: A Highly Efficient Machine Learning-based Ozone Parameterization for Climate Models, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-9358, https://doi.org/10.5194/egusphere-egu25-9358, 2025.
A commonly observed issue in general circulation models is biases in the frequency distribution of precipitation, including too much weak rain (the drizzle problem) and either too much or too little heavy precipitation. High resolution models perform better on this front, but are restricted in the spatial and temporal scales they can simulate.
Previous work (Yuval, O'Gorman, Hill (2021)) demonstrated that training a neural network parameterisation on high-resolution convection-resolving simulations and deploying it within the same model running at lower horizontal resolution can maintain a good representation of precipitation.
Our work builds on this seeking to redeploy the parameterisation within a global atmospheric model, the Community Atmosphere Model (CAM), as a deep convection scheme, with the aim of running stable simulations with improved precipitation prediction. To do so requires interfacing the scheme to operate on a different vertical grid using a different system of variables to the original model in which it was trained.
In this talk we will present this work discussing the objectives alongside the challenges faced moving the parameterisation from one model to another. We share the results from validation in single-column mode against field campaign observations, and of running the scheme globally in an aquaplanet configuration. We will also discuss software architecture and engineering considerations when seeking to develop and redeploy portable machine-learnt parameterisation schemes.
How to cite: Atkinson, J., O'Gorman, P., Berner, J., and Weinzierl, M.: Coupling a new convection parameterisation trained using high-resolution simulations to the Community Atmospheric Model, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-19177, https://doi.org/10.5194/egusphere-egu25-19177, 2025.
Earth System Models (ESMs) are fundamental to understanding and projecting climate change. While they have demonstrated continuous improvements over the last decades, systematic errors and large uncertainties in their projections remain. A large contribution to these uncertainties stems from the representation of unresolved processes such as clouds and convection that occur at scales smaller than the model grid spacing. This impacts the models’ ability to accurately project global and regional climate change, climate variability, and extremes. High-resolution models with horizontal grid spacing of a few kilometers or less alleviate many biases of coarse-resolution models, but at high computational costs. Yet short simulations from high-resolution models can be used to inform machine learning (ML)-based parameterizations that are then incorporated into hybrid (physics+ML) ESMs. This new generation of hybrid models promises to reduce systematic errors and enhance projection capabilities compared to current state-of-the-art ESMs [1, 2]. In an effort to design a comprehensive hybrid ESM, the ICOsahedral Non-hydrostatic (ICON) model is equipped with a variety of physics-aware ML parameterizations, including moist convection, cloud cover and radiation. This talk will present an overview of the modelling activities undertaken within this framework, with a special focus on the developed ML-based cloud cover parameterization. This parameterization takes the form of an interpretable non-linear equation discovered through a combination of ML techniques including symbolic regression and sequential feature selection [3]. We demonstrate that, with this new parameterization, ICON runs stably over several decades and reduces global biases in cloud cover and radiation metrics. In addition, the new equation is controlled by only 10 free parameters that we automatically calibrate to achieve more accurate climate projections. This approach of discovering a low-dimensional data-driven equation for a parameterization with subsequent tuning of the hybrid model can be used in any host ESM provided suitable training data.
References:
[1] Eyring, V., Collins, W.D., Gentine, P. et al., Pushing the frontiers in climate modeling and analysis with machine learning, Nat. Climate Change, doi:10.1038/s41558-024-02095-y, 2024.
[2] Eyring, V., Gentine, P., Camps-Valls, G., Lawrence, D.M., and Reichstein, M., AI-empowered Next-generation Multiscale Climate Modeling for Mitigation and Adaptation, Nat. Geosci., doi:10.1038/s41561-024-01527-w, 2024.
[3] Grundner, A., Beucler, T., Gentine, P. and Eyring, V., Data-driven equation discovery of a cloud cover parameterization, J. Adv. Model. Earth Sys., doi:10.1029/2023MS003763, 2024.
How to cite: Savre, J., Schwabe, M., Grundner, A., Hefner, K., Heuer, H., Klamt, J., Pastori, L., Schlund, M., Gentine, P., and Eyring, V.: Towards a prototype hybrid ICON-ML model with physics-aware machine learning parameterizations, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-17999, https://doi.org/10.5194/egusphere-egu25-17999, 2025.
Renewable resources are strongly dependent on local and large-scale weather situations. Skillful subseasonal to seasonal (S2S) forecasts -beyond two weeks and up to two months- can offer significant socioeconomic advantages to the energy sector. In particular, accurate wind speed forecasts result in optimized generation of wind-based electric power. This study aims to enhance wind speed predictions using a diffusion model with classifier-free guidance to downscale S2S forecasts of surface wind speed. We propose DiffScale, a diffusion model that super-resolves spatial information for continuous downscaling factors and lead times. Leveraging weather priors as guidance for the generative process of diffusion models, we adopt the perspective of conditional probabilities on sampling super-resolved S2S forecasts. We aim to directly estimate the density, associated with the target S2S forecasts at different spatial resolutions and lead times without auto-regression or sequence prediction, resulting in an efficient and flexible model. Synthetic experiments were designed to super-resolve wind speed S2S forecasts from the European Center for Medium-Range Weather Forecast (ECMWF) from a coarse resolution to a finer resolution of data from ERA5, which serves as a high-resolution target, derived from reanalysis data. We achieve a significant increase in the quality of predictions, utilizing the proposed diffusion model for continuous downscaling and bias correction of the ECMWF forecasts.
How to cite: Springenberg, M., Otero Felipe, N., and Ma, J.: DiffScale: Towards Continuous Downscaling and Bias Correction in Subseasonal Wind Speed Forecasts, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-1650, https://doi.org/10.5194/egusphere-egu25-1650, 2025.
Analogue has been a widely-used concept in atmospheric science, particularly useful in weather forecasting and climate-related studies. The underlying idea is straightforward. An analogue is identified by determining its level of similarity to a reference weather or climate condition, traditionally, via computing a Euclidean distance. Recently, a deep-learning based framework, called ClimaDist, was proposed for climate analogue identification, found to outperform traditional Euclidean distance metrics. Despite the promising performance, similarly to many deep-learning models, it is challenging to estimate the uncertainty of the analogue searching process undertaken by ClimaDist. This hinders its applicability to real-world operations, especially for those requiring decision making.
To address this challenge, this study extends the capabilities of ClimaDist through incorporating a uncertainty quantification method, together with explainable AI (XAI) techniques. Specifically, the Evidential Deep Learning (EDL) approach is applied to the analogue searching process undertaken by the ClimaDist. This enables effective quantification of the uncertainty associated with data and model, respectively, while exploring their relationship with overall model performance. Two distinct scenarios are applied to these two models using data that were seen and unseen during the training processing.
An experiment has been designed to verify the proposed approach using ERA5 data over a square domain centred at the Nettebach (Germany) covering the geographic range of 55°N to 47°N and 3°E to 11°E. Two ClimaDist models, one with the best validation performance and the other one best training performance, respectively, are used for comparison. These models are assessed based on the similarity of the found analogues and via under two distinct scenarios –with input data seen and unseen during the training process, respectively. Preliminary results suggest that the integration of uncertainty quantification enhances the interpretability and reliability of analogue identification, enabling improved downstream applications. Specifically, high model uncertainty can be highlighted by the proposed approach while fully unseen data is used as input. This not only provides valuable insight in knowing the capacity of the underlying model but also allows the optimization of resource usage.
How to cite: Chen, C., Chen, P.-C., Tseng, C.-Y., and Wang, L.-P.: Uncertainty quantification through the climate analogue identification process by ClimaDist, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-3610, https://doi.org/10.5194/egusphere-egu25-3610, 2025.
Generative Deep Learning architectures, such as Diffusion models, offer an alternative to traditional physical modeling and regression models due to their ability to produce stochastic ensembles with a single run. Even though these models are capable of downscaling coarse data, they are often trained in contained regions, which can lead to severe spatial overfitting as the model learns location-specific patterns rather than generalizable physical relationships. In practice, the usability of the models is constrained to the area where they were originally trained, and their predictive capabilities degrade significantly when applied to regions outside the training domain, even if these regions share similar characteristics.
This study presents a one-step and two-step diffusion model capable of downscaling 2-meter temperature from ERA5 to higher-resolution grids in large areas, such as the Contiguous United States or Europe, without spatially overfitting. We use CONUS404, a reanalysis dataset created using simulations of the Weather Research and Forecasting (WRF) model over the Contiguous United States, as our target data and ERA5 and constants involved in the creation of CONUS404, such as altitude and land use, as our input. The model has been trained over the whole area using 10 years of 3-hourly data, and two years have been used for testing. To study the spatial generalization capabilities of the model, we reserve an area of the study region solely for testing and compute evaluation metrics separately for this area to ensure meaningful results. We compare the results of training in large and small areas and the number of years. In addition, we discuss the usefulness of ensemble prediction and the effect that the number of ensemble members has on the performance of the downscaling. Future steps include applying this methodology for downscaling EURO-CORDEX to EMO1 and multivariate downscaling.
How to cite: Benitez Benavides, M., Rodríguez, M., Margalef, T., Panadero, J., and Dutta, O.: Stochastic diffusion model for large-scale temperature downscaling, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-19397, https://doi.org/10.5194/egusphere-egu25-19397, 2025.
In hydrology, the use of machine learning (ML) has gained traction due to its ability to provide alternative or complementary approaches to traditional process-based modelling. These models identify numerical patterns in time series data without needing to solve conservation equations. This flexibility enables hydrological calculations in areas where data sources are incomplete or non-existent.
Studies benchmarking ML models (SVM, RNN, CNN) against process-based models have shown that ML models deliver promising results with lower computational cost and less information about the physical processes they are modelling. Consequently, they can effectively utilize spatially discretized physical data on a large scale.
This study designs a Long Short-Term Memory (LSTM) neural network to learn sequential relationships between atmospheric, climatic and geographic features and daily streamflow data from 39 headwater gauging stations in the northern Ebro river basin. LSTM models include an internal state that can store information and learn long-term dependencies, enabling them to model sequential data effectively. However, the numerical patterns identified by LSTM models do not inherently respect universal physical laws, such as the conservation of mass.
To address the limitation, the Mass-Conserving LSTM (MC-LSTM) model has been employed and compared with the standard LSTM model. The MC-LSTM model introduces a modified cell structure that adheres to conservation laws by extending the learning bias to model the redistribution of mass.
This analysis highlights not only the high accuracy of LSTM models in predictive hydrologic modelling but also the critical importance of integrating physics-based features to enable ML models to effectively capture the hydrological dynamics of the basin.
Acknowledgments: This work is funded by the European Research Council (ERC) through the Horizon Europe 2021 Starting Grant program under REA grant agreement number 101039181-SED@HEAD.
How to cite: González Planet, I. and Juez, C.: Streamflow forecasting in the Ebro river basin using Machine Learning (ML) and a physical mass constraint, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-8105, https://doi.org/10.5194/egusphere-egu25-8105, 2025.
Local climate information is crucial for impact assessment and decision-making, yet coarse global climate simulations cannot capture small-scale phenomena. Current statistical downscaling methods infer these phenomena as temporally decoupled spatial patches. However, to preserve physical properties, estimating spatio-temporally coherent high-resolution weather dynamics for multiple variables across long time horizons is crucial. We present a novel generative approach that uses a score-based diffusion model trained on high-resolution reanalysis data to capture the statistical properties of local weather dynamics. After training, we condition on coarse climate model data to generate weather patterns consistent with the aggregate information. As this inference task is inherently uncertain, we leverage the probabilistic nature of diffusion models and sample multiple trajectories. We evaluate our approach with high-resolution reanalysis information before applying it to the climate model downscaling task. We then demonstrate that the model generates spatially and temporally coherent weather dynamics that align with global climate output.
How to cite: Schmidt, J., Schmidt, L., Strnad, F., Ludwig, N., and Hennig, P.: Spatiotemporally Coherent Probabilistic Generation of Weather from Climate, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-11083, https://doi.org/10.5194/egusphere-egu25-11083, 2025.
The performance of a deep convolutional neural network in predicting near-surface air temperature (T2m) and total precipitation (P) over Europe is assessed, comparing its results with the Copernicus European Regional Reanalysis (CERRA) and the regional dynamical model HCLIM. The ML-model accurately captures broad seasonal temperature and precipitation patterns, with minor biases in summer and more pronounced warm biases in winter. While the model effectively reproduces the probability density functions (PDFs) of daily temperature and precipitation, it underestimates extreme cold events and the high precipitation extremes in some regions. Climate indices, including cold extremes (TM2PCTL), warm extremes (TM98PCTL), consecutive dry days (CDD), and consecutive wet days (CWD), highlight that the ML model aligns closely with CERRA. However, it slightly underestimates CDD and overestimates CWD, particularly in mountainous and Mediterranean regions. Analyses of spatio-temporal variability demonstrate high correlations with CERRA for temperature, exceeding 0.99 for spatial patterns and 0.95 for temporal correlations, while correlations for precipitation are lower, with underestimated temporal variability. The ML model generally outperforms HCLIM, particularly in aligning with observed data, although challenges remain in capturing extremes and reducing biases in certain regions. These results further highlight the potential of the ML model for regional climate downscaling and impact studies, while emphasizing the need for further refinement to enhance its representation of extreme events and improve spatial accuracy.
How to cite: Fuentes-Franco, R., Ivanov, M., Koenigk, T., Krus, K., Aldama Campino, A., and Wang, F.: Assessment of a Pan-European high-resolution downscaling through Deep Learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-17278, https://doi.org/10.5194/egusphere-egu25-17278, 2025.
High-resolution climate projections are vital for understanding the local impacts of climate change in fields like agriculture, hydrology, energy production, and disaster risk management. However, Earth System Model (ESM) output often lacks the spatial detail needed to capture regional to local-scale variability, while showing large biases when compared to observational data. Statistical downscaling (SD) is commonly used to address such issues by refining the coarse spatial resolution of ESM output. While the current ISIMIP3 (Inter-Sectoral Impact Model Intercomparison Project, third round) SD algorithm is robust and computationally efficient, it struggles with increasing differences between source and target resolutions. To address these limitations, we applied deep-learning-based SD methods to create a globally consistent, high-resolution dataset for near-surface climate variables.
Using the perfect prognosis approach, we combined ERA5 as large-scale atmospheric predictors with ERA5-Land as high-resolution predictands (target resolution of ~10 km) to create accurate transfer functions (TFs) that align with ISIMIP's requirements, such as trend preservation and inter-variable consistency. These TFs are subsequently applied to ESM output to generate downscaled climate forcings. The resulting framework is both scalable and computationally efficient, making it suitable for multi-model applications. The results were compared with similar methodologies and its improvements were demonstrated in a cross-validation framework, particularly in capturing local-scale features.
Our approach offers a robust tool for generating high-resolution climate data, providing valuable insights to researchers and decision-makers working on climate impact assessments and adaptation planning. This work contributes to the next iteration of ISIMIP and to OptimESM, targeting the CMIP7-based modeling framework. The derived high-resolution projections are designed to complement CMIP7 datasets, enabling the creation of downscaled ensembles that conform with ISIMIP's objectives and support a wide range of impact modeling applications.
How to cite: Quesada-Chacón, D., Sauer, I., Mengel, M., and Frieler, K.: A deep learning approach to statistical downscaling and its potential to increase the resolution of the impact model simulations within ISIMIP, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-12175, https://doi.org/10.5194/egusphere-egu25-12175, 2025.
Estimation of the Reference Evapotranspiration (ET0) is critical in water resources management under climate change, especially for agricultural water management in arid and semi-arid regions. Thus, estimating baseline ET0 poses significant challenges, particularly in inadequate climatological monitoring regions. In this study, a hybrid modeling approach based on the incorporation of empirical models, Particle Swarm Optimization (PSO), and XGBoost algorithm (Empirical-PSO-XGBoost) was developed and evaluated to forecast ET0 under limited climate variables. The results showed the Empirical-PSO-XGBoost outperformed the purely calibrated empirical and Temperature-PSO-XGBoost models for estimating monthly (daily) ET0 with NSE reaching 0.99 (0.86) and 0.98 (0.67) for the calibration and validation phases, respectively. Besides, up to 63 CMIP6 projections were coupled with Empirical-PSO-XGBoost for forecasting the long-term ET0 under SSP245 and SSP585 climate change scenarios. Thus, the simulation showed a significant increase in ET0 and seasonal patterns compared to the baseline ET0 where the change in range of [+5, +10] % is associated with probability values of 0.65 and 0.78 for SSP245 and SSP585, respectively. Overall, the developed framework is useful for implementing adaptation strategies to mitigate climate change effects on water resource allocation and agricultural management. It provides the ET0 associated with Exceedance probability for each month which is useful for assessing the water availability-related-risk in scheduling irrigation and sowing date of crops.
How to cite: Elbilali, A., Hadri, A., Taleb, A., EL Khalki, E. M., Tanarhte, M., and Kharrou, M. H.: A Novel Modeling Framework based on Empirical models, PSO, XGBoost, and multiple GCMs for the projection of Long-Term Reference Evapotranspiration, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-12646, https://doi.org/10.5194/egusphere-egu25-12646, 2025.
Climate variability and change significantly influence crop production, presenting challenges that extend from understanding the basic crop growth principles to evaluating the effects of extreme weather events on crop development. Addressing this requires effective agro-management strategies guided by tailored climate services. However, a critical gap exists between scientific insights and their practical application. This study introduces and evaluates an AI-driven methodology designed to simulate crop growth and predict grain maize yields across Europe. Specifically, nested Recurrent Neural Networks (RNNs) are tested as a computationally efficient surrogate model for the process-based ECroPS model developed by the European Commission’s Joint Research Centre. Traditional mechanistic crop models, like ECroPS, require numerous meteorological inputs and significant computational resources, limiting scalability for applications such as large-scale climate simulations or ensemble modeling that explore variables like climate projections and CO₂ effects. In contrast, the surrogate AI model relies on just three weather inputs—daily minimum and maximum temperatures and daily precipitation—trained using ECMWF-ERA5 reanalysis data. This streamlined approach demonstrates the potential to bridge the gap between resource-intensive crop modeling and scalable, data-driven solutions for climate impact assessments.
How to cite: Vlachopoulos, O., Luther, N., Ceglar, A., Toreti, A., and Xoplaki, E.: Surrogate impact modelling for crop yield assessment with nested RNNs, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13061, https://doi.org/10.5194/egusphere-egu25-13061, 2025.
Despite the remarkable strides made by AI-driven models in modern precipitation forecasting, these black-box models cannot inherently deepen the comprehension of underlying mechanisms. To address this limitation, we propose an AI-driven knowledge discovery framework known as genetic algorithm-geographic weighted regression. Through this framework, we have constructed an iterative optimization of knowledge generation and utilization. On the one hand, new explicit equations are discovered to describe the intricate relationship between precipitation patterns and terrain characteristics. Experiments have shown that the discovered equations demonstrate remarkable accuracy when applied to precipitation data, outperforming conventional empirical models. Notably, our research reveals that the parameters within these equations are dynamic, adapting to evolving climate patterns. On the other hand, these previously undisclosed equations have contributed new knowledge about terrain-precipitation relationships, which can be embedded into the AI model for better interpretability and climate projection accuracy. Specifically, the unveiled equations can enable fine-scale downscaling for precipitation predictions using low-resolution future climate data. This capability offers invaluable insights into the anticipated changes in precipitation patterns across diverse terrains under future climate scenarios, which enhances our ability to address the challenges posed by contemporary climate science.
How to cite: Xu, H., Chen, Y., Zeng, Z., Li, N., Li, J., and Zhang, D.: Exploring Terrain-Precipitation Relationships with Interpretable AI for Advancing Future Climate Projections, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-17628, https://doi.org/10.5194/egusphere-egu25-17628, 2025.
Accurate climate projections are critical for understanding climate change and to design adaptation and mitigation strategies. Weighting schemes that aggregate a range of climate model projections are widely used to provide more reliable estimates of future climate conditions. Recently, causal discovery has been successfully introduced in the weighting schemes to constrain uncertainties in climate model projections based on the performance and interdependence of climate models. However, the previous methodologies typically (and strongly) only utilize a single metric, the F1 score of performance and similarity between each climate model and observational data, to compare the different models' causal structures. Here, we introduce alternative and more sophisticated causal weighting schemes inspired by the theory of kernel methods and Gaussian processes to compare causal graphs directly in suitable reproducing kernel Hilbert spaces. In addition, we propose alternative causal weighting schemes that rely on interventions, graph-based distances, and counterfactual evaluations. We will evaluate the causal weighting strategies in various synthetic and CMIP6 model datasets.
How to cite: Camps-Valls, G., Debeire, K., Varando, G., Runge, J., and Eyring, V.: Causal Weighting for Climate Projections, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-8363, https://doi.org/10.5194/egusphere-egu25-8363, 2025.
Understanding the complex relationship between trace gases as well as undestanding various source and sink pathways in the atmsophere need good qualtity continuous and reliable datasets. However, obtaining comprehensive long-term profiles for key trace gases is a significant challenge. We have initiated a new research strand to consrtuct long term data using machine learning. Output from a Chemical Transport Model (CTM) and observational data from satellite instruments (such as HALOE and ACE-FTS) is merged using machine learning. This integration results in the creation of daily, gap-free datasets for six crucial gases: ozone (O3), methane (CH4), hydrogen fluoride (HF), water vapour (H2O), hydrogen chloride (HCl), and nitrous oxide (N2O) from 1991 to 2021.
Chlorofluorocarbons (CFCs) are a critical source of chlorine that controls stratospheric ozone losses. Currently, ACE-FTS is the only instrument that provides sparse but daily measurements of these gases. Monitoring changes in these ozone-depleting substances, which are now banned, helps assess the effectiveness of the Montreal Protocol. We have initiated the construction of gap-free stratospheric profile data for CFC-11 as a subsequent step.
We use a regression model to estimate the relationship between various tracers in a CTM and the differences between the CTM output field and the observations, assuming all errors are due to the CTM setup. Once the regression model is trained for observational collocations, it is used to estimate biases for all the CTM grid points. To enhance accuracy, we employed various regression models and found that XGBoost regression outperforms other methods. ACE-FTS v5.2 data (2004-present) is used to train (70%) and test (30%) the XGBoost performance.
Our results demonstrate excellent agreement between the constructed profiles and satellite measurement-based datasets. Biases in TCOM data sets, when compared to evaluation profiles, are consistently below 10% for mid-high latitudes and 50% for the low latitudes, across the stratosphere. The constructed daily zonal mean profile datasets, spanning altitudes from 15 to 60 km (or pressure levels from 300 to 0.1 hPa), are publicly accessible through Zenodo repositories.
CH4: https://doi.org/10.5281/zenodo.7293740
N2O: https://doi.org/10.5281/zenodo.7386001
HCl : https://doi.org/10.5281/zenodo.7608194
HF: https://doi.org/10.5281/zenodo.7607564
O3: https://doi.org/10.5281/zenodo.7833154
H2O: https://doi.org/10.5281/zenodo.7912904
CFC-11: https://doi.org/10.5281/zenodo.11526073
CFC-12: https://doi.org/10.5281/zenodo.12548528
COF2: https://doi.org/10.5281/zenodo.12551268
In an upcoming iteration, we are enhancing the algorithm as well as add more species in the current setup. We believe these data sets would provide valuable insights into the dynamics of stratospheric trace gases, furthering our understanding of their behaviour and impact on the climate.
References:
Dhomse, S. S., et al.,: ML-TOMCAT: machine-learning-based satellite-corrected global stratospheric ozone profile data set from a chemical transport model, Earth Syst. Sci. Data, 13, 5711–5729, https://doi.org/10.5194/essd-13-5711-2021, 2021.
Dhomse, S. S. and Chipperfield, M. P.: Using machine learning to construct TOMCAT model and occultation measurement-based stratospheri
c methane (TCOM-CH4) and nitrous oxide (TCOM-N2O) profile data sets, Earth Syst. Sci. Data, 15, 5105–5120, https://doi.org/10.5194/essd-15-5105-2023, 2023.
How to cite: Dhomse, S. and Chipperfield, M.: Machine Learning to Construct Daily, Gap-Free, Long-Term Stratospheric Trace Gases Data Sets, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-2394, https://doi.org/10.5194/egusphere-egu25-2394, 2025.
Complex multiscale dynamics of the atmosphere in extratropical latitudes includes various persistent atmospheric regimes with the residence time up to several weeks. Identification, simulation and prediction of such dynamics remains one of the challenging problems. In this work we use a stochastic recurrent neural network (RNN) with specific architecture to address this problem, appealing to RNN’s ability to handle memory effects well. The proposed RNN connects two types of variables: (i) a low-dimensional representation of the physical variables via Principal Component Analysis (PCA), and (ii) Kernel PCA variables which serve to better represent the target atmospheric regimes [1]. The stochastic component of the RNN has a simple form which allows us to analytically write Bayesian log-posterior and log-likelihood functions to train and cross-validate the model given the particular dataset.
Using the observed and climate-model-generated winter geopotential height data in the Northern Hemisphere, we show that the proposed stochastic model is able to reproduce/predict various dynamical properties and distributions of the target regimes in the kernel space, as well as to reconstruct kernel variables from a low-dimensional representation of the original spatio-temporal field.
References
1. Mukhin et al. (2022). Revealing recurrent regimes of mid-latitude atmospheric variability using novel machine learning method. Chaos: An Interdisciplinary Journal of Nonlinear Science, 32(11).
How to cite: Gavrilov, A., Mukhin, D., Safonov, S., and Samoilov, R.: Stochastic recurrent neural network for modeling atmospheric regimes, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7404, https://doi.org/10.5194/egusphere-egu25-7404, 2025.
Recent years have seen rapid advancements in large-scale data-driven models for weather forecasting. Several of these models can now compete with, and in some respects outperform, physics-based numerical models for medium-range forecasting. They offer significant computational savings and potential forecasting accuracy improvements approximately equivalent to a decade of progress in traditional methods. This progress has prompted announcements from weather institutes across the world about plans to integrate AI-driven models into their operational workflows in the near future.
As data-driven models become integral to operational forecasting, critical questions about fairness and equity remain. Studies reveal substantial variations in forecast quality across regions, particularly for extreme weather. Unlike physical models, the disparities in data-driven models often stem from passive design decisions, such as inductive biases and weighting schemes, which may be reassessed and changed, if needed. Moreover, ensuring equitable access to these models, along with the means to effectively utilise and improve them, is essential so that both high- and low-income countries can share in their benefits.
This work explores fairness in data-driven weather forecasting, with a focus on outcome-based perspectives. We begin by defining fairness from both process and outcome viewpoints. We then analyse the performance of current data-driven models across different regions and socio-economic groups globally. Our findings reveal significant disparities that may exacerbate pre-existing socio-economic and climate-related vulnerabilities. To address these challenges, we advocate for a deliberate focus on fairness and equity in data-driven model development, emphasising the importance of active design choices to promote equitable outcomes.
How to cite: Olivetti, L. and Messori, G.: Whose weather is it? A fairness perspective on data-driven weather forecasting, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-11668, https://doi.org/10.5194/egusphere-egu25-11668, 2025.
XAI finds signs of clouds in the Net Ecosystem Exchange of boreal forest
Laanti, Lintunen, Noe, Miles, Heljanko, Kulmala, Ezhova
We applied three distinct machine learning models (random forest, LightGBM, and XGBoost) to predict net ecosystem exchange (NEE) in boreal forests using site-level information and climatic variables from two Finnish stations, SMEAR I and II as well as one Estonian station, SMEAR Estonia. Our study focuses on explainable artificial intelligence (XAI) technique called Shapley values, to interpret how radiation and meteorological and biospheric variables influence NEE.
Using XAI, we found that diffuse radiation enhancement of NEE is linked to type of cloudiness. Our Shapley value analysis revealed that at the same diffuse radiation level, NEE can be enhanced more under overcast sky than under clear-sky or broken cloudiness conditions. Under a certain parameter range, this seems to counterbalance the negative effect of reduction in PAR on photosynthesis under overcast sky. Furthermore, visualizing the interplay between PAR, cloudiness, and NEE based on seasonality highlighted subtle differences in how these parameters interact at northern versus southern sites. Importantly, the use of three distinct machine learning models that all showed similar results demonstrate that these observed relationships are consistent.
Although the discovered relationships between radiation, cloudiness and NEE do not necessarily reflect true causality, they can guide further testing of possible causal hypotheses. By integrating XAI into NEE modeling with machine learning, we gain deeper insights into the physical and ecological processes shaping carbon fluxes. Such interpretability is vital for understanding NEE dynamics in boreal forests, particularly in the face of evolving climate scenarios where cloud cover, temperature, and moisture regimes shift and introduce complex feedback mechanisms. Integrating XAI thus provides a valuable framework for interpreting complex, potentially nonlinear drivers behind NEE and for exploring new avenues of causal investigation in ecosystem research.
How to cite: Laanti, T., Ezhova, E., Lintunen, A., Noe, S., Kulmala, M., Miles, V., and Heljanko, K.: XAI finds signs of clouds in the Net Ecosystem Exchange of boreal forest , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13234, https://doi.org/10.5194/egusphere-egu25-13234, 2025.
Understanding the potential impacts of climate change on global vegetation dynamics is crucial for effective environmental management and biodiversity conservation. This study employs a machine learning-based framework to analyze historical NDVI data and project future vegetation growth under different climate scenarios. Utilizing the GIMMS NDVI dataset (1981–2000) for model training and CMIP6 climate projections (2021–2100) for scenario analysis, the study evaluates changes in vegetation growth across four Shared Socioeconomic Pathways (SSPs). Results indicate a significant near-term increase in global mean NDVI (2021–2040) under all scenarios, followed by divergent trends. While SSP126 and SSP245 sustain modest increases, SSP370 and SSP585 show sharp declines in NDVI over the long term, driven by adverse temperature effects. Regional analyses reveal contrasting patterns: NDVI values in Africa, South America, and Oceania decline under most scenarios, while North America, Europe, and Asia exhibit potential increases, except under high-emission scenarios like SSP585. These findings underscore the importance of targeted interventions to mitigate climate impacts and highlight the role of machine learning in predicting vegetation responses to environmental changes. The study provides actionable insights for policymakers, emphasizing the need for sustainable land management practices and greenhouse gas reduction strategies to preserve global ecosystems.
How to cite: Nguyen, A. K. and Chen, W.: Machine Learning Projections of Climate Change Impacts on Global Vegetation Dynamics, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15622, https://doi.org/10.5194/egusphere-egu25-15622, 2025.
Earth’s seasonality profoundly influences nearly every aspect of life on our planet. It plays a key role in driving vegetation cycles and shaping wildlife behavior. Seasonality also impacts human life significantly, affecting health, mood, social dynamics, and cultural patterns. Despite its importance, seasonality is still traditionally defined by astronomical seasons—equal-length divisions applied uniformly across the Earth. Although this division is simple and intuitive, it overlooks crucial seasonal patterns influenced by atmospheric weather. In this study, we propose a data-driven approach to redefining seasons using objective clustering. We develop an algorithm that segments various meteorological factors into meaningful seasonal clusters. Building on this algorithm, we objectively define seasons for each region globally and analyze the effects of Climate Change on these clusters. We find that seasonality is driven by different meteorological factors in different regions on Earth. Additionally, we observe that Climate Change has significantly altered the duration and onset of Earth’s seasons.
How to cite: Shmuel, A., Magaritz-Ronen, L., Raveh-Rubin, S., and Milo, R.: Revisiting Earth’s Seasonality using Machine Learning Models, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-16234, https://doi.org/10.5194/egusphere-egu25-16234, 2025.
Posters virtual: Fri, 2 May, 14:00–15:45 | vPoster spot 2
EGU25-15382 | Posters virtual | VPS30
Event-Based Physics-Informed Neural Networks for Dust Storm PredictionFri, 02 May, 14:00–15:45 (CEST) | vP2.1
Dust storms pose significant environmental challenges in arid and semi-arid regions, causing serious health, environmental, and socio-economic impacts. Traditional dust modeling approaches, like numerical methods, often struggle to balance accuracy, computational efficiency, and data availability. This study employed a Physics-Informed Neural Network (PINN) model for event-based dust storm modeling, integrating the physical principles of dust dynamics with data-driven methods. We demonstrated the applicability of the above framework in the Lake Urmia Basin, where the lake desiccation and external dust sources have triggered local dust storms. To this end, we first analyzed ground-recorded PM10 and weather data to identify dusty days between 2004 and 2019. Next, we trained an initial neural network (NN) model with remote sensing data that describe meteorological and boundary layer characteristics at the locations of pollution monitoring stations. This approach allowed us to generate gridded PM10 data, overcoming the limitations posed by insufficient and non-continuous data for directly training PINN. Finally, the PINN model was trained and validated on 21 selected dust events from three stations chosen for their spatial distribution and sufficient availability of PM10 data throughout the events. Analysis revealed that the initial NN model achieved R² of 62% and mean absolute error (MAE) of 65 on the test data. The PINN model demonstrated substantial improvement with mean R² of 93% and mean MAE of 9 on the gridded PM10, and MAE of 39 when validated against ground observations. Furthermore, the model yielded lower prediction accuracy in urban compared to rural stations, which is attributed to the bias imposed by the influence of terrestrial and industrial pollutions. This study demonstrates the effectiveness of PINNs in tackling dust transport modeling challenges in data-sparse regions, providing a novel way to combine physical principles with data-driven techniques for large-scale environmental applications.
How to cite: giahchin, K. and danesh-yazdi, M.: Event-Based Physics-Informed Neural Networks for Dust Storm Prediction, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15382, https://doi.org/10.5194/egusphere-egu25-15382, 2025.
