In recent years, increasing attention has been given to hybrid methods that combine machine learning with physical models. These approaches are now being applied across planetary and heliophysical domains, from forecasting solar eruptions and solar wind conditions, to automating the analysis of planetary surfaces or improving on-board data handling. They demonstrate how data-driven methods can benefit from physical knowledge, while physics-based models can be improved through modern data analysis techniques.
This session aims to provide an inclusive and interdisciplinary forum for researchers applying machine learning in planetary sciences and heliophysics, as well as those developing methods at the intersection between data-driven and physics-based approaches. We particularly encourage contributions that illustrate the wide range of applications, encourage exchange between disciplines and showcase the transition from research to operations.
Unlike Earth, Mars does not possess an upstream solar wind monitor. This lack of continuous solar wind observations has fundamentally limited scientific studies that investigate solar wind impacts on the Mars space environment, and with increasing relevance, operational tasks for predicting space weather at the planet. Previous estimates of the solar wind have been pursued through physics-based modeling (e.g. magnetohydrodynamic models) or empirical (e.g. assuming statistical relationships with downstream observations) proxies. Proxies are often based on downstream observations from multiple orbiting spacecraft. These spacecraft pass in and out of the bow shock providing a semiregular sampling of the pristine solar wind. The most complete, and ongoing, set of the solar wind’s magnetic field and plasma parameters is from the NASA MAVEN spacecraft. MAVEN has orbited Mars since 2014, but additional assets add resolution to this dataset such as including ESA’s MEX mission which has been in orbit since 2003, the CNSA’s Tianwen-1 orbiter since 2021, and NASA’s ESCAPADE mission scheduled for orbital insertion in 2027.
In this presentation we will summarize a prior effort to create a continuous solar wind estimation upstream from Mars. This virtual solar wind monitor, or vSWIM (see Azari, Abrahams, Sapienza, Halekas, Biersteker, Mitchell, Pérez et al., 2024, doi: 10.1029/2024JH000155) was trained and assessed on MAVEN data with Gaussian process regression. Gaussian process regression, a type of machine learning, was used to provide predictions, and uncertainties on these predictions, at various temporal resolutions. vSWIM currently enables informed solar wind estimation at Mars for most of the time since 2014. We will then discuss current progress on improving vSWIM’s capacity for multi-spacecraft integration for enhanced operational space weather prediction efforts at Mars.
How to cite: Azari, A., Hayes, K., and Rutala, M.: Probabilistic Solar Wind Estimation for Operational Space Weather Prediction at Mars, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6050, https://doi.org/10.5194/egusphere-egu26-6050, 2026.
The properties and the spatial distribution of the large-scale structures of the Solar Corona (SC) determine the observed solar wind structure at 1 AU. Coronal Holes (CHs) are the primary source of the fast solar wind, which is the most geoeffective component of solar wind, and they appear as large dark patches in the Extreme Ultraviolet (EUV) images from the Atmospheric Imaging Assembly (AIA) on the Solar Dynamic Observatory (SDO) and the Extreme Ultraviolet Imaging Telescope (EIT) on the Solar and Heliospheric Observatory (SoHO). These observatories provide images of the SC at different wavelengths, which enables the identification of CH morphology and other large-scale structures along a given line of sight. It is crucial to understand the CH regions and their properties for effective space weather forecasting. This work is part of the WindTRUST project, with the primary goal of improving the reliability of solar wind models for space weather forecasting. Here, we aim to develop an automatic threshold-based CH detection tool for predictions across solar cycles 23, 24, and 25. We also plan to integrate this CH detection tool into a solar wind model validation pipeline, creating a fully automated validation system that provides a quantitative assessment of predictions. We categorized the large-scale features of the SC, such as active regions, solar flares, coronal mass ejections (CMEs), and filaments, based on their spatial distribution, phase of the solar cycle, and additional properties, including the GOES solar flare class. A Sequential Neural Network (NN) model was then trained by optimizing the architecture of the hidden layers to achieve higher predictive accuracy. The resulting model estimates the threshold required for integration into the Coronal Hole (CH) detection scheme, thereby enabling automated, consistent identification of CH boundaries in EUV images across solar cycles 23, 24, and 25. To interpret the performance of our NN model, we divided the predicted CH results into solar minimum and maximum cases across the solar cycles 23, 24, and 25. We also provide a comparison of our CH detection results with those obtained from other detection tools. Once we identify CH contours from our model, we validate them using a diagnostic test against CH contours from the Potential Field Source Surface (PFSS) model (non-MHD) and the WindPredict (WP) model (Polytropic and Alfven Wave) (MHD). Finally, we couple the CH detection tool with the validation pipeline to develop an automation tool for solar wind predictions.
How to cite: Henadhira Arachchige, K. H. P., Perri, B., Brun, A.-S., Strugarek, A., Buchlin, E., Reville, V., and Ausseresse, M.: AI-Based Coronal Hole Detection and Solar Wind Model Validation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5189, https://doi.org/10.5194/egusphere-egu26-5189, 2026.
Automated identification of coronal structures using machine-learning techniques can support forecasting of extreme solar events, enable autonomous solar-observing missions, and accelerate understanding of physical processes in the solar atmosphere. Existing approaches typically focus on large-scale regions or adopt conservative segmentation strategies that limit structural detail. We train a lightweight variant of the You-Only-Look-Once (YOLO) object-detection framework [1] and, in parallel, design a scheme based on classical computer-vision operations and morphological filtering. Both are compared against the deep-learning-based SCSS-Net [2]. All three frameworks detect active regions and coronal holes in images from the Atmospheric Imaging Assembly onboard the Solar Dynamics Observatory. To reduce bias, training and testing use masks from multiple sources, including SPoCA [3], CHIMERA [4], Region Growth [5], and custom annotations. Methods are evaluated for scientific performance and computational cost using standard metrics such as the Dice score and Intersection over Union (IoU). We further assess on-board feasibility by outlining potential use cases and current technical limitations, and by evaluating performance on raw, uncalibrated data to ensure operational compatibility and robustness. Finally, we examine coronal hole mapping across multiple AIA wavelength channels and analyse correlations with signed and unsigned magnetic flux.
References
[1] Redmon et al. "You only look once: Unified, real-time object detection." Proceedings of the IEEE conference on computer vision and pattern recognition. 2016.
[2] Mackovjak et al. "SCSS-Net: solar corona structures segmentation by deep learning." Monthly Notices of the Royal Astronomical Society 508.3 (2021): 3111-3124.
[3] Verbeeck et al. "The SPoCA-suite: Software for extraction, characterization, and tracking of active regions and coronal holes on EUV images." Astronomy & Astrophysics 561 (2014): A29.
[4] Garto et al. "Automated coronal hole identification via multi-thermal intensity segmentation." Journal of Space Weather and Space Climate 8 (2018): A02.
[5] Tlatov, A., K. Tavastsherna, and V. Vasil’eva. "Coronal holes in solar cycles 21 to 23." Solar Physics 289.4 (2014): 1349-1358.
How to cite: Gonidakis, P., Poedts, S., and Magdalenic, J.: Machine Learning for Solar Coronal Structure Segmentation on SDO AIA Data and Applications, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11886, https://doi.org/10.5194/egusphere-egu26-11886, 2026.
Reliable solar flare forecasting is limited by two forms of class imbalance in active region time series: (i) the overwhelming dominance of the non-flaring, quiet state over the eruptive state, and (ii) the insufficient separability between common, physically similar event classes (e.g. C-class versus M-class flares). Although empirical parameters derived from the photospheric vector magnetic field (VMF), such as those provided by SDO/HMI SHARP products, capture aspects of active region complexity and free energy buildup, they often evolve smoothly and overlap across flare classes. Consequently, while many models can distinguish between flares and no-flares reasonably well, they struggle to distinguish flare magnitude and association with eruptive phenomena (e.g. CMEs) using photospheric information alone. This suggests that improved flare-class separation requires (a) the explicit definition of what constitutes 'similarity' between pre-flare states, and (b) parametrization that emphasizes flare-relevant structure over common active region features.
We investigate a representation learning strategy that combines the parametrization of SDO/HMI SHARP VMF cutouts using a Variational Autoencoder (VAE) with a contrastive stage to reshape the resulting embedding geometry. First, a VAE is trained to encode SHARP cutouts into compact latent vectors that capture active region morphology. These vectors are then refined using a Siamese-like objective constructed from weak supervision, which uses event labels and empirical SHARP parameters as proxies for elevated flare likelihood. The contrastive stage then uses this weak supervision to encourage a latent geometry that better reflects flare-relevant evolution. This study emphasizes latent-space structure, i.e. neighborhood consistency and class-conditional clustering, and evaluates whether these properties facilitate improved probabilistic prediction across multiple forecast horizons, by training lightweight downstream models on (i) empirical parameters, (ii) VAE latents and (iii) their combined representations.
How to cite: Dineva, E., Magdalenic, J., Miloshevich, G., Gonidakis, P., Carella, F., and Poedts, S.: Probabilistic Solar Flare Forecasting via Weakly Supervised Contrastive Refinement of VAE Latent Spaces, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18188, https://doi.org/10.5194/egusphere-egu26-18188, 2026.
Both observations and simulations have revealed that magnetic reconnection occurs at thin current sheets within the transition region of collisionless shock waves. These ion- and electron-scale structures arise from stream instabilities and turbulence in the shock layer, contribute significantly to repartition of energy across the shock, and propagate far into the downstream region. In a recent study [Gingell et al. 2023, Physics of Plasmas, 30, 0123902], a series of 2D hybrid particle-in-cell simulations were used to explore the shock-driven generation and decay of reconnecting structures over a broad range of parameters. Magnetic field line integration was used to quantify reconnection in each simulation, classifying each cell in the domain as having “closed” or “open” magnetic field topology. Here, we use these classifications to train a convolution neural network (CNN) to identify regions of the simulation that are undergoing (or have undergone) magnetic reconnection. This is performed by splitting each simulation domain into a series of 1D virtual trajectories, with a view to creating a dataset equivalent to a series of in situ observations. We find that the trained CNN is able to effectively identify structures of interest in simulations that have different plasma and shock parameters to the training data set, as well as in those with different dimensionality (i.e. 3D simulations). Further, we present a pipeline for applying this simulation-trained CNN to in situ observations of shocks by the Magnetospheric Multiscale and Solar Orbiter spacecraft, and demonstrate successful detection of reconnection sites embedded in the shock layer. We discuss these techniques more generally as a case study for using machine learning to identify structures of interest in spacecraft data, which may contribute to on-board event selection for burst modes in spacecraft with relatively limited downlink capacity.
How to cite: Gingell, I. L.: Connecting hybrid plasma simulations of collisionless shockwaves to in situ observations with machine learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3944, https://doi.org/10.5194/egusphere-egu26-3944, 2026.
In our solar system, the main source of plasma is the Sun, which produces the so-called solar wind by continuously pushing its outermost layer -the corona- into space. The turbulent solar wind impinges on our planet and interacts with its magnetic field, creating a region of space called Earth’s magnetosphere. From its birth to its impact on our planet, the solar wind still harbors numerous unanswered questions. Answering these questions requires the numerical modelling of the plasma itself.
The most physically accurate numerical methods are based on kinetic modeling, which tracks the particles' velocity distribution function. However, these methods are numerically demanding since they involve modeling the complex six-dimensional particle distribution function as it evolves in time. To simplify the problem, such distribution is integrated over the velocity coordinates leading to the (more efficient) three-dimensional fluid plasma framework. Still, the passage to the fluid equations comes with an important caveat. The fluid system of equations needs to be closed by choosing a proper “closure”. The objective of this work is to tackle the closure problem by employing a combination of kinetic simulation and machine learning techniques.
We perform multiple decaying turbulence plasma simulations using a Hybrid-PIC [1] (i.e. kinetic ions, fluid electrons) model. By varying different physical parameters, notably the ion beta, we explore the variability of the solar wind. These kinetic simulations serve as the ground truth to train a machine learning model. The machine's task is to "learn" the best approximation for the closure equation. We focus in particular on the reconstruction of the pressure tensor. We explore various machine learning techniques [2, 3] (CNN, GAN, FNO) that have shown promise in atmospheric science but are new to this specific problem. We show how this reconstructed closure performs better than other analytical approximations [4] (polytropic, CGL, CGL+FLR effects). The final goal is to learn a closure equation that can effectively incorporate complex kinetic physics into a simplified, yet more accurate, fluid simulation. This will significantly increase the fidelity of solar wind models without making them prohibitively expensive to compute.
[1] Behar, Etienne, Shahab Fatemi, Pierre Henri, e Mats Holmström. «Menura: A Code for Simulating the Interaction between a Turbulent Solar Wind and Solar System Bodies». Annales Geophysicae 40, fasc. 3 (2022): 281–97. https://doi.org/10.5194/angeo-40-281-2022.
[2] Kovachki, Nikola, Zongyi Li, Burigede Liu, et al. «Neural Operator: Learning Maps Between Function Spaces». Preprint, 2 maggio 2024. https://doi.org/10.5555/3648699.3648788.
[3] Jeong, Hyun-Jin, Mingyu Jeon, Daeil Kim, et al. «Prediction of the Next Solar Rotation Synoptic Maps Using an Artificial Intelligence–Based Surface Flux Transport Model». The Astrophysical Journal Supplement Series 278, fasc. 1 (2025): 5. https://doi.org/10.3847/1538-4365/adc447.
[4] Hunana, P., A. Tenerani, G. P. Zank, et al. «An Introductory Guide to Fluid Models with Anisotropic Temperatures. Part 1. CGL Description and Collisionless Fluid Hierarchy». Journal of Plasma Physics 85, fasc. 6 (2019): 205850602. https://doi.org/10.1017/S0022377819000801.
How to cite: Dazzi, P., de Oliveira Lopes, F. N., Jeong, H.-J., Calvet, E., and Keppens, R.: Modelling of space plasma from Vlasov to fluid: machine learning applied to the closure problem, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17564, https://doi.org/10.5194/egusphere-egu26-17564, 2026.
Accurate prediction of rare but high-impact events is a recurring challenge in planetary science and heliophysics, where strongly imbalanced data distributions are common (e.g. extreme space-weather conditions). Standard empirical risk minimization tends to bias machine-learning models toward frequently observed regimes, often leading to poor performance on scientifically and operationally critical tail events. Existing mitigation strategies, such as loss re-weighting or synthetic over-sampling, have shown mixed and problem-dependent success.
We present PARIS (Pruning Algorithm via the Representer theorem for Imbalanced Scenarios), a data-centric framework that addresses imbalance by optimizing the training dataset itself rather than modifying the loss function or model architecture. PARIS exploits the representer theorem for neural networks to compute a closed-form representer deletion residual, which quantifies the change in validation loss induced by removing an individual training sample—without requiring retraining. Using an efficient Cholesky rank-one downdating scheme, this enables fast, iterative pruning of uninformative or performance-degrading samples.
We demonstrate PARIS on a real-world space-weather regression problem (Dst prediction), where it reduces the training set by up to 75% while preserving or improving overall RMSE and outperforming loss re-weighting, synthetic over-sampling, and boosting baselines. These results highlight representer-guided dataset pruning as a computationally efficient, interpretable, and physically relevant approach for improving rare-event regression in heliophysics and related planetary science applications.
Preprint: https://www.arxiv.org/abs/2512.06950
How to cite: Camporeale, E.: PARIS: Pruning Algorithm via the Representer theorem for Imbalanced Scenarios, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6702, https://doi.org/10.5194/egusphere-egu26-6702, 2026.
In GNSS-denied deep space exploration missions, high-precision state estimation and navigation positioning are critical to ensuring the successful completion of complex mission objectives. However, the environmental characteristics of extraterrestrial surfaces, such as drastic illumination changes, monotonous textures, and sparse features, often lead to the failure of traditional visual navigation systems. Meanwhile, IMUs, despite their high-frequency and anti-interference capabilities, face the challenge of integration error accumulation caused by biases and noise. Although Visual-Inertial Odometry (VIO) achieves complementary advantages through multi-source fusion, existing end-to-end deep learning methods often lack explicit physical modeling, this deficiency leads to a sharp degradation in generalization performance and susceptibility to drift in extreme environments, thereby failing to meet the stringent standards required for aerospace-grade missions.
To address the extreme environments of extraterrestrial bodies and the limitations of existing methods regarding the lack of physical consistency and insufficient generalization, we propose a Physics-Aware Hybrid Deep Visual-Inertial Odometry (PDVIO) navigation method suitable for extraterrestrial bodies, this framework is dedicated to deeply coupling physics-driven kinematic priors with data-driven deep representation capabilities to construct a navigation system that possesses both strong robustness and high precision. Specifically, this study comprises three core contributions: First, addressing the integration drift caused by IMU noise, we designed an analytical physical pre-integration module based on Lie Group Theory, unlike traditional networks that directly regress pose parameters, this module explicitly constructs IMU motion differential equations on the SE(3) manifold, embedding hard rigid body dynamic constraints directly into the network structure, thereby substantially reducing the risk of model divergence in extreme environments. Second, to cope with visual perception degradation caused by high-dynamic illumination changes and sparse textures, we introduce a FlowNet-enhanced multi-scale feature encoder, by extracting hierarchical spatiotemporal optical flow features via a pyramid structure, this enables the system to effectively capture ego-motion states based on optical flow field consistency even in regions with extreme textures, significantly enhancing the stability of front-end tracking. Finally, addressing the drawback of traditional methods relying on fixed noise covariance, we propose a differentiable factor graph back-end framework based on Graph Attention Networks (GAT). Utilizing an attention mechanism to dynamically learn the confidence weights of visual and inertial modalities according to the real-time dynamic environment, this successfully achieves adaptive end-to-end joint optimization from feature extraction to state estimation, greatly improving the system's adaptability and navigation accuracy in complex deep space environments.
Experiments conducted on simulation datasets and real-world ground data demonstrate that, while maintaining the efficiency of deep learning feature extraction, this method significantly enhances the robustness and generalization capability of the navigation system, specifically, the trajectory estimation error is markedly reduced compared to traditional end-to-end models, effectively mitigating long-term integration drift. Therefore, this study not only validates the effectiveness of embedding physical priors into deep learning frameworks, addressing the issues of insufficient robustness and limited autonomy inherent in purely data-driven methods within aerospace scenarios, but also provides a highly reliable and high-precision navigation solution for future planetary exploration missions involving precise pinpointing and navigation.
How to cite: Jiao, Y., Liu, S., Xiao, C., Ouyang, W., and Tong, X.: Physics-Aware Hybrid Deep Visual-Inertial Odometry Based on Graph Attention Networks for GNSS-denied Environment, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2564, https://doi.org/10.5194/egusphere-egu26-2564, 2026.
Modeling the lunar terrain is a key challenge for lunar missions, having impact on mission planning, resource planning and the establishment of sustainable human bases on the Moon. Thanks to the Lunar Reconnaissance Orbiter (LRO) and its Narrow Acquisition Cameras (NACs) acquiring images at a spatial resolution of 50cm, a large collection of images are now available. Despite this, automatically generating digital elevation models (DEMs) on the Moon remains a challenge. Classic methods like multi-view stereovision or photoinclinometry struggle with lunar specificities such as the large shadows and the permanently shadowed regions (PSRs) and the absence of atmosphere, the complex lighting conditions and the homogeneity of the lunar surface texture.
In 2020, a new self-supervised neural-network-based method called Neural Radiance Fields (NeRF) was introduced and demonstrated outstanding 3D reconstruction capacities from multi-view images. Recent advancements adapted the methodology to the challenging field of satellite imagery of the Earth and exhibited competing or even better results than classic methodologies. Some recent works tried to transfer to the Moon but either constrained their studies to simulated data or rather reused existing models.
In this work, we explore the potential of NeRF to learn the 3D shape of the lunar surface at a very high resolution from LRO NACs data, supported by a coarse estimation of the ground given by processed data from LRO’s altimetric sensor called the Lunar Orbiter Laser Altimetry (LOLA). Our main contributions are the generation of a LRO NeRF-ready dataset on a Moon South Pole region that we intend to openly share and the development of a specific model coined LuNeRF. We demonstrate that, with an adapted radiance modeling, LuNeRF can recover the geometry of small craters, as well as perform novel view synthesis and relighting tasks.
How to cite: Thenoz, C., Derksen, D., Malapert, J.-C., and Schmidt, F.: LuNeRF: How Neural Radiance Fields Can Advance Very High Resolution Lunar Terrain Reconstruction, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6977, https://doi.org/10.5194/egusphere-egu26-6977, 2026.
A specific aspect of space weather that remains poorly understood is the exchange of information from space to the ground through the ionosphere. A central component of this process involves understanding how current systems such as field-aligned currents transfer energy and momentum between the magnetosphere and the ionosphere. However, the non-linear behaviour of these current systems poses significant challenges for identifying the drivers of ionospheric currents and understanding the inner dynamics of the ionosphere itself.
To tackle these complexities and their effects on the ground, we adopt a data-driven approach using space-based observations from the Active Magnetosphere and Planetary Electrodynamics Response Experiment (AMPERE). Specifically, we focus on gaining insights into what drives these current systems by finding underlying statistical patterns (dominant modes) in the data using unsupervised machine learning methods. We employ techniques such as β - Variational Autoencoders (β-VAEs), which have been proven useful in identifying patterns in unlabelled observational data.
We extract dominant modes and connect them to physical drivers of the system with a two-step approach. First, we quantify model performance using a physically motivated goodness-of-fit metric to ensure that the learned model reconstructions capture the essential dynamics in the current system. Second, we analyse the model’s latent space, representing a compressed representation of the high dimensional input data. We then analyse the latent space and connect the influence of the individual latents to physical drivers of the system through the usage of the OMNI dataset. This approach enables a systematic interpretation of the model’s internal representations in terms of underlying physical processes.
How to cite: Hasler, M., Coxon, J., and Smith, A.: AI-driven analysis of dangerous space weather: finding dominant modes in space-based measurements, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2757, https://doi.org/10.5194/egusphere-egu26-2757, 2026.
Recent studies in machine learning (ML) and geology have demonstrated a strong potential for automated classification of rocks and minerals. Though, the performance of ML models like Convolutional Neural Networks (CNNs), for pattern recognition of geological textures remains limited under controlled microscopic imaging conditions. This study explores the possibility of automated classification of multiple rocks and minerals including visually similar samples using microscopic texture information.
Initially, microscopic images of terrestrial basalt and magnetite which are visually similar under RGB microscopy, were captured using a digital USB microscope under varying illumination and magnification settings. These materials were selected to evaluate the performance of CNN models on differences in grain size, crystallinity and surface reflectance. A dataset comprising 2500 images per class was created and expanded using several augmentation techniques to increase the robustness of the model. With transfer learning, multiple models were trained amongst which InceptionV3 model achieved the highest validation accuracy for the initial binary classification problem.
The trained model achieved a validation accuracy of 98.30% and a test accuracy of 95%, demonstrating strong generalization capabilities. To assess the model’s effectiveness, performance metrics such as Precision, F1-Score, Confusion Matrix and ROC curve were examined. These findings provide insight into the strengths of CNN based pattern recognition in geological applications and demonstrate how deep learning techniques can be used for automated texture based classification.
Also, while this study does not directly utilize planetary datasets, it establishes a foundation for future applications of texture based ML methods in autonomous rover operations for geological analysis. We aim to extend this study to multiple basaltic variants and lithological classes under conditions relevant to Martian exploration, for building robust ML algorithms which can be used for geological image analysis.
How to cite: Shrivastava, S., Rema, A., Kumar, S., and Bhatia, M. P. S.: Texture based classification of geological materials using Deep Learning - Proof of concept for Planetary Surface Analysis, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4379, https://doi.org/10.5194/egusphere-egu26-4379, 2026.
In many plasmas, physical processes of relevance occur over ranges of scales covering many orders of magnitude. Thus, modelling plasmas comes with a trade-off between physical accuracy and computational cost. Fully kinetic models correctly self-consistently describe collisionless plasmas by advancing the velocity distribution functions (VDFs) in time, either directly (Vlasov methods) or sampling it through computational particle (PIC codes). A computationally cheaper but physically less accurate alternative are multi-fluid models. Instead of the VDFs, these models evolve fluid quantities and can approximate kinetic processes of interest by choosing a suitable closure for the hierarchy of fluid moment equations, i. e., an equation for the divergence of the heat flux in the case of ten-moment fluid models. In most heliospheric plasmas, including for example the solar wind, the observed VDFs are non-Maxwellian, which gives rise to many different instabilities that exchange energy between particles and fields. We investigate the use of machine learning models for the discovery of heat flux closures, as an alternative to the typically employed Hammett-Perkins-like analytical closures. As a test case, we use the two-stream instability, which occurs when there is a large velocity drift between two electron populations with respect to their thermal speed, and causes the formation of electron holes and electric field saturation in its nonlinear stage. While the linear stage of the two stream instability is well reproduced by 10-moment models with analytical closures, reproducing electric field evolution at saturation is a challenge for reduced models. In this work, we compare fully kinetic Vlasov simulations against two-fluid 10-moment simulations employing both analytical and ML-driven closures.
How to cite: Jeß, E., Lautenbach, S., Köhne, S., Grauer, R., and Innocenti, M. E.: Comparing analytical and machine learning heat flux closures, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6301, https://doi.org/10.5194/egusphere-egu26-6301, 2026.
Accurate forecasting of extreme solar flares is essential for mitigating space weather impacts on critical infrastructure, yet current deep learning approaches face fundamental limitations in operational reliability. Models often lack physical interpretability and may fail to generalize to configurations under-represented in training data, which are critical weaknesses
when forecasting rare extreme events. We take steps toward addressing these gaps by developing physics-informed architectures that embed magnetohydrodynamic (MHD) constraints directly into neural network training.
Using SDO/HMI SHARP vector magnetograms (2010–2021, 13,298 observations), we compare three approaches for 24-hour multi-class flare forecasting: (1) a ResNet34 baseline, (2) a reconstruction-physics hybrid enforcing MHD constraints through magnetic field reconstruction, and (3) a probability-physics hybrid coupling physics-derived features to classification probabilities. The probability-physics model achieves macro-averaged True Skill Statistic (TSS) of 0.389 [95% CI: 0.355–0.425] versus baseline 0.338 [0.301–0.375], a statistically significant 15% improvement (p < 0.001). Critically, physics-constrained models reduce divergence violations by two orders of magnitude, ensuring predictions satisfy fundamental conservation laws and remain physically interpretable across a broader range of magnetic configurations, including those under-represented in training data.
Feature space analysis reveals that intermediate C-class flares occupy transitional magnetic states with extensive overlap between non-flaring and extreme configurations, highlighting an intrinsic forecasting challenge that persists across architectures. M+ (M- and X-class) events maintain strong discrimination (AUC > 0.87) despite severe class imbalance, indicating that physically meaningful features can aid identification of extreme events even when training samples are scarce.
Our results suggest that embedding first-principles MHD constraints—divergence-free conditions, force-free equilibrium, and energy conservation—enhances both forecast skill and physical plausibility without increasing computational cost. The integration of physics-informed learning with CNN-based flare prediction offers a pathway toward improving operationally deployed systems with enhanced reliability for extreme event forecasting. For operational forecasters, improved physical interpretability may provide greater confidence in model predictions during critical decision-making, while reduced false alarm rates minimize unnecessary protective actions for satellite operators and power grid managers.
Keywords: extreme space weather, solar flare forecasting, physics-informed neural net-
works, operational reliability, magnetohydrodynamics, infrastructure risk mitigation
How to cite: Bristol-Alagbariya, A., Eastwood, J., and Moseley, B.: Integrating Physics-Informed Neural Networks with Convolutional Neural Networks for Solar Flare Prediction, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9809, https://doi.org/10.5194/egusphere-egu26-9809, 2026.
How to cite: Tao, S., Turc, L., Dahani, S., Lipsanen, V., Kalliokoski, M., Ojuva, M., Aunai, N., Zhang, H., Wang, S., and Raptis, S.: Automated Identification of Foreshock Transients, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16681, https://doi.org/10.5194/egusphere-egu26-16681, 2026.
Solar flares are intense eruptive events caused by the rapid release of magnetic energy, often impacting Earth's space environment through electromagnetic radiation and high-energy particles. Accurate flare prediction is critical for space weather forecasting. However, many existing deep learning approaches often rely on single-modal inputs or shallow feature fusion, limiting their ability to capture complementary information. In this study, we propose a dual-branch multimodal fusion deep learning model for 24-hour solar flare prediction. The model integrates magnetograms and magnetic parameters through cross-attention mechanisms, followed by cross-scale interactions at the feature level to enhance multi-scale representation. It is designed to perform both binary prediction of ≥ C-class flares and multi-class classification of C, M, and X-class flares. To ensure rigorous evaluation, we employ a stratified group five-fold cross-validation scheme to preserve class representativeness and adopt a splitting-before-sampling strategy based on active region number to prevent data leakage. Experimental results show that the model achieves a TSS of 0.661 and an HSS of 0.630 for binary ≥ C-class prediction, while notably attaining a TSS of 0.780 and an HSS of 0.785 for X-class flares in the multi-class task. Compared with existing approaches, the model demonstrates superior performance in predicting intense X-class flares, effectively suppresses the false alarm rate, and exhibits strong generalization capability.
How to cite: Limin, Z., Xingyao, C., Xiaoshuai, Z., Dong, Z., and Yihua, Y.: The Deep Learning-Based Dual-Branch Multimodal Fusion Model for Solar Flare Prediction, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16741, https://doi.org/10.5194/egusphere-egu26-16741, 2026.
Sub-Alfvénic solar wind intervals predominantly transpire into the core of magnetic clouds (MC) during interplanetary coronal mass ejection (ICME) events, facilitating an intense mode of solar wind-magnetosphere interaction wherein energy and information can propagate via magnetic field lines. These phenomena are associated with intense magnetic fields, low plasma beta, heightened Alfvénic activity, and exceptionally effective energy transfer to the magnetospheric domain. This study employs a physics-informed machine learning framework to identify and characterize the sub-Alfvénic magnetic cloud regime using data from many solar cycles. A feature space motivated by physical principles is established based on the plasma characteristics of upstream solar wind observed from the L1 point, along with metrics of wave activity obtained from time-frequency analysis. Employing unsupervised machine learning, the high-dimensional solar-wind feature space is mapped onto a low-dimensional latent space that elucidates the intrinsic organization of solar-wind plasma regimes. By integrating recognized MC occurrences and disparate individual case studies of sub-Alfvénic flow onto the established phase-space map, it has been deduced that severe coupling conditions are indicative of a cohesive global regime of solar wind behavior rather than isolated anomalies. This framework also illustrates transition paths among background solar wind, sheaths, and magnetic cloud cores, utilizing the evolution of coupling conditions during interplanetary coronal mass ejection passages.
How to cite: Haldar, S.: A Data-Driven Phase-Space View of Sub-Alfvénic Magnetic-Cloud Coupling, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18327, https://doi.org/10.5194/egusphere-egu26-18327, 2026.
Current analyses of solar differential emission measure predominantly rely on two-dimensional (2D) imaging and interpretation, which inherently limit our ability to fully capture the true three-dimensional (3D) characteristics of coronal structures and dynamic processes. This 2D perspective consequently hinders a comprehensive understanding of the complex physical processes governing the solar atmosphere.
To address these limitations, we present a novel methodology for the spatio-temporal reconstruction of the low solar corona, with several machine learning techniques. This approach enables us to reconstruct several physical parameters, including EUV radiation, temperature, and electron density, across varying altitudes and observation time. Based on these 3D reconstruction results, our method can further generate synthetic observational images from various viewpoints and times, providing a comprehensive visualisation of the corona's dynamic 3D structure. Furthermore, it can estimate missing wavelength observations for missions such as Solar Orbiter. This significantly supports multi-spacecraft collaborative observations and data fusion efforts. Besides, our reconstructed results can also serve as an enhanced initial state for coronal and interplanetary simulations.
How to cite: Liu, J., Poedts, S., Shen, C., and Liu, J.: Automatic Spatio-Temporal Differential Emission Reconstruction Method, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18421, https://doi.org/10.5194/egusphere-egu26-18421, 2026.
Radio occultation (RO) techniques provide valuable remote-sensing insights into planetary ionospheres and atmospheres by measuring the bending of radio signals as they traverse atmospheric layers. Mutual radio occultations between the Trace Gas Orbiter (TGO) and Mars Express (MEX) demonstrated the feasibility of this approach but were limited by hardware not designed for radio science occultation measurements—most notably, the absence of ultra-stable oscillators, single-frequency operation, and restricted timing precision.
The Mars Magnetosphere ATmosphere Ionosphere and Space-weather SciencE (M-MATISSE) mission—currently in its Phase A study by the European Space Agency (ESA)—is a Medium-class (M7) candidate that will overcome these constraints through the dedicated MaCro (Mars Crosslink Radio Occultation) instrument: a dual-frequency, precision-timed, ultra-stable radio system purpose-built for inter-satellite occultations. MaCro’s design enables high-accuracy profiling of the Martian ionosphere and atmosphere across diverse geometries and solar conditions.
This study systematically investigates how the known limitations of TGO–MEX influenced the retrieved electron density profiles and explores how modern machine-learning techniques—for example regression-based drift correction—can enhance the data-processing pipeline. The outcomes of this work will support the development of MaCro’s data processing chain and contribute to the improvement of its performance.
How to cite: Andert, T., Pätzold, M., Vorderobermeier, T., Hahn, M., Tellmann, S., Oschlisniok, J., Peter, K., and Haser, B.: Advancing Inter-Satellite Radio Occultation with MaCro on the M-MATISSE Mission, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18949, https://doi.org/10.5194/egusphere-egu26-18949, 2026.
In this work, we aim to advance the prediction of solar wind speed several days in advance. The approach is based on analyzing solar coronal images in conjunction with solar wind speed. We create labelled data pairs from over a decade of EUV images obtained from the SDO/AIA and solar wind data at 1AU recorded by ACE, WIND, and DISCOVR. We use the archived SDO machine-learning ready dataset (SDO-ML), and the solar wind speed at 1AU from the NASA OMNIWEB dataset. We construct a deep neural network model and capture the temporal component of the solar wind propagation with a time-dependent neural network, e.g., Recurrent Neural Network. Physical constraints are incorporated to train the model and optimize the prediction. The generalization capability of our model is investigated via cross-validation, whereby careful separation into training, validation, and test datasets is performed as a function of solar activity. We report on the impact of the deep neural network architecture as a universal function approximation in its ability to capture the temporal relationship between solar EUV characteristics and solar wind speed at 1 AU.
How to cite: Delouille, V., Li, K., Kamalabadi, F., and Davila, J.: Physics-informed time-dependent deep neural network for solar wind prediction, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-22082, https://doi.org/10.5194/egusphere-egu26-22082, 2026.
Posters virtual: Wed, 6 May, 14:00–18:00 | vPoster spot 1b
EGU26-11945 | Posters virtual | VPS22
SEPNET: a multi-task deep learning framework for SEP forecastingWed, 06 May, 14:06–14:09 (CEST) vPoster spot 1b
Solar phenomena such as flares, coronal mass ejections (CMEs), and solar energetic particles (SEPs) are actively monitored and assessed for space weather hazards. In recent years, machine learning has demonstrated considerable success in solar flare forecasting. Accurate SEP forecasting remains challenging in space weather monitoring due to the complexity of SEP event origins and propagation. We introduce SEPNET, an innovative multi-task neural network that integrates forecasting of solar flares and CME summary statistics into the SEP prediction model, leveraging their shared dependence on space-weather HMI active region patches (SHARP) magnetic field parameters. SEPNET incorporates long short-term memory and transformer architectures to capture contextual dependencies. The performance of SEPNET is evaluated on the state-of-the-art SEPVAL SEP dataset and compared with classical machine learning methods and current state-of-the-art pre-eruptive SEP prediction models. The results show that SEPNET achieves higher detection rates and skill scores while being suitable for real-time space weather alert operations.
How to cite: Chen, Y., Yu, Y., Zhao, L., Whitman, K., Manchester, W., and Gombosi, T.: SEPNET: a multi-task deep learning framework for SEP forecasting, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11945, https://doi.org/10.5194/egusphere-egu26-11945, 2026.
