SWR6 – Applications of Artificial Intelligence to Space Weather and Space Climate

Talks

SWR6.1 Tue 5/11 09:00-10:15, room C2A – Mondego

Author(s): michele piana, sabrina guastavino, francesco marchetti, anna maria massone, paolo massa, anna volpara

MIDA, dipartimento di matematica università di genova; MIDA, dipartimento di matematica università di genova; dipartimento di matematica, università di padova; MIDA, dipartimento di matematica, università di genova; fachhochschule nordwestschweiz, switzerland; MIDA, dipartimento di matematica, università di genova

Abstract: Solar flares are the most explosive events in the heliosphere and are recognized as the main trigger of space weather. The two main open research issues concerning solar flares are: 1) are the information contained in the magnetic fields constrained  into solar active regions accurate enough to allow flare forecasting? and, 2) which are the physical mechanisms that explain energy release and electron acceleration in coronal plasma before and during flares’ occurrence? This talk proposes to discuss some hints concerning the solution to these two questions and show some results obtained by applying inversion methods, and machine and deep learning to different kinds of remote sensing data.

Author(s): George Miloshevich, Francesco Carella, Panagiotis Gonidakis, Alessandro Bemporad, Giovanni Lapenta, Stefaan Poedts

KU Leuven; KU Leuven; KU Leuven; INAF-Osservatorio Astrofisico di Torino; KU Leuven; KU Leuven

Abstract: Predicting solar flares has important applications in communications, safety of operations for satellites, etc., which is why it is relevant in the context of space weather. Most of the previous work using Machine Learning (ML) recently focused on classifying the flares, assigning labels such as M, X, etc. Such discrete labels may overlook physical information rather than the underlying X-ray flux, which is a continuous label.
We use Convolutional Neural Networks (CNNs) to predict X-ray flux from a combination of Helioseismic and Magnetic Imager (HMI) and Extreme Ultraviolet (EUV) images of the Sun using a curated ML dataset [1] created by Solar Dynamics Observatory (SDO) AIA and HMI instrument data. The inputs are chosen to represent the relevant layers of the solar atmosphere: the line-of-sight magnetograms from HMI (photosphere), 94 Å (flaring regions), 171 Å (quiet sun), 193 Å (coronal structures), 304 Å (chromosphere) AIA wavelengths.  Data are processed to have time matching between SDO images [1] and GOES X-ray fluxes 30 minutes maximum. Limb-brightening correction using a geometrical function is performed on the relevant AIA wavelengths to avoid possible biases in the predictions. In addition, we compare and contrast the usefulness of original full disk images vs the synoptic maps as the inputs to the CNN.
We evaluate the benefit of introducing soft constraints to predict the extremes better. We utilize the self-supervised framework known as Model Genesis [2], originally developed for medical imaging. In this approach, the CNN is pre-trained using an encoder-decoder architecture designed to reconstruct original images of the sun after applying artificial transformations. These transformations include non-rigid deformations and local pixel shuffling. By initializing the CNN with weights from this reconstruction task, we enable it to learn robust features that improve performance on downstream tasks. Next, the encoder head is attached to the CNN to make predictions about the X-ray flux.  In order to compare our work to other classifications, we post-process the outputs to better associate the predictions of the X-ray flux with the flare indices. Our goal is to benchmark this approach with state-of-the-art approaches using skills such as True Skill Score (TSS) for categorical predictions and Brier Skill Score (BSS) for probabilistic prediction on a hold-out test set. Finally, using eXplainable AI (XAI) [3], one can find statistical estimates of which active regions and parts of the solar atmosphere may contribute the most to the prediction of flares. [1] Galvez, R., et al. The Astrophysical Journal Supplement Series 242.1 (2019): 7.[2] Zhou, Z., et al. Medical image analysis 67 (2021): 101840.[3] Letzgus, S. et al. (2022). IEEE Signal Processing Magazine, 39(4), 40-58.

Author(s): Shane Maloney, Paul Wright, Sophie Muarry, Peter Gallagher, Anna Massone, Michele Piana, Sabrina Guastavino, Edoardo Legnaro, Tamas Kiss, Gabor Terstyanszky, Dimitris Kagialis, Robert Lovas, Attila Farkas

Dublin Institute for Advanced Studies; Dublin Institute for Advanced Studies; Dublin Institute for Advanced Studies; Dublin Institute for Advanced Studies; Università di Genova; Università di Genova; Università di Genova; Università di Genova; University of Westminster; University of Westminster; University of Westminster; SZTAKI Institute for Computer Science and Control; SZTAKI Institute for Computer Science and Control

Abstract: The Active Region Classification and Flare Forecasting (ARCAFF) project aims to deliver cutting-edge solar flare predictions via SolarMonitor.org, a familiar and user-friendly interface. Leveraging rich solar flare and active region classification datasets from NOAA SWPC and the Met Office, a key objective of ARCAFF is to generate near-real-time active region detection and classifications using full-disk magnetograms. These automated classifications will complement our curated flare lists to (1) feed existing statistical models based on the flaring rates of individual active region classes and (2) be the basis of end-to-end deep learning flare forecasting pipelines.
We will discuss how these hand-labelled datasets have been curated and evaluated and the methodologies used to develop explainable deep learning models for active region classification. Additionally, we will explore the transition from research to operations, emphasising the role of manual classifications provided by NOAA SWPC and the Met Office in benchmarking and validating the performance of these models in operational settings. This continual monitoring–enabled by expert forecasters–is crucial in ensuring the accuracy and reliability of active region classification and space weather prediction facilitated by ARCAFF.

Author(s): Alan Wood, Yaqi Jin, Luca Spogli, Jaroslav Urbar, Eelco Doornbos, Gareth Dorrian, Daria Kotova, Rayan Imam, Lucilla Alfonsi, Kasper van Dam, Mainul Hoque, Wojciech Miloch

University of Birmingham, UK; University of Oslo, Norway; Istituto Nazionale di Geofisica e Vulcanologia, Italy; Institute of Atmospheric Physics CAS, Prague, Czech Republic; The Royal Netherlands Meteorological Institute (KNMI), the Netherlands; University of Birmingham, UK; University of Oslo, Norway; Istituto Nazionale di Geofisica e Vulcanologia, Italy; Istituto Nazionale di Geofisica e Vulcanologia, Italy; The Royal Netherlands Meteorological Institute (KNMI), the Netherlands; German Aerospace Center (DLR), Germany; University of Oslo, Norway

Abstract: Machine learning and statistical modelling are powerful tools which can be used in combination with large datasets to build models of complex systems. The resulting models have the potential to make predictions about the future behaviour of such systems. If appropriate proxies can be found for the driving processes, then these proxies can be included in the models as explanatory variables and the influence of the driving processes on the quantity which is being predicted can be estimated.
However, many machine learning techniques suffer from a major limitation. It is clear which explanatory variables have been selected for inclusion within a model, but not why these choices have been made. In recent years physics-informed machine learning has been used to address this limitation, by using prior knowledge of the underlying physical system to guide the modelling process.
In this work, we present an alternative approach to address this limitation, by combining two modelling methods. Generalised Linear Modelling (GLM; McCullagh and Nelder, 1983) is a statistical modelling method which has been used since the 1970s in fields including medical trials, road safety and agriculture. In recent years it has been applied to build predictive models of ionospheric variability (i.e. Wood et al., 2024; Spogli et al., 2024). This technique has distinct advantages. It is computationally cheap and it needs input from the modeller at every step of the process, so it is clear why particular terms have been included in the model. We use a two-step modelling method. The first step uses GLM to determine which terms should be included in the model and the second step uses an artificial neural network to optimise the model.
Models are created of the plasma density, the ionospheric variability on horizontal spatial scales between 100 km and 20 km and the thermospheric density in Low Earth Orbit (LEO) using observations from the European Space Agency (ESA) Swarm mission. The performance of these models is assessed using measures of accuracy, bias, precision and association, and the improvement in the model performance of using our two-step modelling method, compared to only using GLM or a neural network, is determined.
Finally, the steps required to turn these results into operational models of plasma density, plasma variability and the thermospheric density in LEO are discussed.
This work is within the framework of the Swarm Space Weather: Variability, Irregularities, and Predictive capabilities for the Dynamic ionosphere (Swarm-VIP-Dynamic) project, funded by ESA in the “4D IONOSPHERE – EXPRO+” framework (ESA Contract No. 4000143413/23/I-EB).
References
McCullagh, P. and Nelder, J. A. 1983. Generalized linear models. CRC monographs on statistics and applied probability, Chapman and Hall, London. ISBN 10:0412238500.
Spogli, L. et al., (2024). Statistical Models of the Variability of Plasma in the Topside Ionosphere: 2: Performance Assessment, J. Space Weather Space Clim., swsc230023. https://doi.org/10.1051/swsc/2024003
Wood., A. G. et al., (2024). Statistical models of the variability of plasma in the topside ionosphere: 1. Development and optimisation, J. Space Weather Space Clim., 14, 7, DOI: https://doi.org/10.1051/swsc/2024002

SWR6.2 Tue 5/11 14:15-15:15, room C2A – Mondego

Author(s): Alexa Halford, Kyle Murphy, Vivan Liu, Katherine Garcia-Sage, Joshua Pettit, Jonathon Smith, Jeffery Klenzing

NASA GSFC; Lakehead University/Independent Researcher; John Hopkins University; NASA/GSFC; Catholic University/GSFC; Catholic University/GSFC; NASA.GSFC

Abstract: Atmospheric density is key in determining the drag experienced by low-Earth orbiting satellites. During elevated solar and geomagnetic activity periods, enhanced atmospheric density drives increased satellite drag, degrading orbits, leading to large uncertainties in orbit determination and propagation. In the most extreme cases, this degradation in orbit can reduce satellite lifetimes. As such, quantifying the dynamics of atmospheric density is of great importance to satellite operators and in space weather modeling and forecasting. Of particular importance is understanding the dynamics of atmospheric density through geomagnetic storms, during which density rapidly varies on timescales of a few hours and by nearly an order magnitude; a combination of solar and geomagnetic effects drives these variations. In this talk, we investigate atmospheric density’s solar and geomagnetic drivers during geomagnetic storms (storm time) and during geomagnetic quiet periods (quiet time) using a Random Forest machine learning algorithm to model atmospheric density. Three models are developed using different sets of independent variables: one using daily solar indices (typically used in atmospheric modeling); a second using the high cadence Flare Irradiance Spectral Model 2 (FISM2) data set; and a third driven by a combination of the FISM2 data set and geomagnetic indices. During quiet times, all three models perform well; however, during geomagnetic storms, the combined FISM2/geomagnetic indices model performs significantly better than the models based solely on solar activity. Further study then considers the relative importance of the impacts of different drivers of more impulsive space weather; for example, solar flares, magnetospheric compression, and storm main phase dynamics. Overall, this work demonstrates the importance of including geomagnetic activity in modeling atmospheric density and serves as a proof of concept for using machine learning algorithms to model and, in the future, forecast atmospheric density for operational use.

Author(s): Matthew Billcliff, Andy Smith, Wai Lok Woo, Jonathan Rae, Matt Owens, Luke Barnard, Nathaniel Edward-Inatimi

Northumbria University; Northumbria University; Northumbria University; Northumbria University; University of Reading; University of Reading; University of Reading

Abstract: Geomagnetic storms cause significant disruptions in the magnetosphere. Large storms can disrupt satellite operations, communication systems, and power grids, causing significant technological and economic impacts. Current storm forecasting models most often utilise data from L1 satellites upstream of the Earth, constraining our lead time to a few hours, often insufficient for effective mitigation measures. Enhancing the lead time for forecasting geomagnetic disturbances is crucial for protecting infrastructure and ensuring operational continuity.
To extend lead times, we must utilise solar data. Greater lead times come with greater uncertainties about which structures will hit earth, which we can capture with an ensemble. In this work, we propagate solar wind to Earth using an ensemble of the HUXt numerical model. The HUXt numerical model uses a one-dimensional upwind method, as opposed to a full 3D heliospheric MHD model, allowing faster computation times, and hence a larger number of ensemble members compared to MHD based models. The output of each ensemble member is used as input to time series classifier machine learning models. A final model then aggregates these models by training on output of the base models. Our target variable for quantifying the disturbance of a storm is the Hpo index, a modern version of the Kp index, offering a better distribution and cadence, for assessing geomagnetic activity.
In this work, we evaluate multiple model architectures for time series classification, testing their performance across various storm intensities. We implement a variety of informative metrics for assessing the model performance, including RMSE, Brier Skill Score, and dynamic time warping as a method of counteracting timing uncertainties in the data.
Overall, we show that the coupled numerical model and machine learning approach from solar data improves our lead time on Hpo forecasting.

Author(s): Fabiana Camattari, Sabrina Guastavino, Emma Perracchione, Katsiaryna Bahamazava, Daniele Telloni, Michele Piana, Anna Maria Massone

MIDA, Dipartimento di Matematica, Università degli Studi di Genova, Genova (Italy); MIDA, Dipartimento di Matematica, Università degli Studi di Genova, Genova (Italy); Dipartimento di Scienze Matematiche Giuseppe Luigi Lagrange, Politecnico di Torino, Torino (Italy); Dipartimento di Scienze Matematiche Giuseppe Luigi Lagrange, Politecnico di Torino, Torino (Italy); Istituto Nazionale di Astrofisica (INAF), Osservatorio Astrofisico di Torino, Pino Torinese (Torino, Italy); MIDA, Dipartimento di Matematica, Università degli Studi di Genova, Genova (Italy); MIDA, Dipartimento di Matematica, Università degli Studi di Genova, Genova (Italy)

Abstract: In this study we deal with the challenging problem of forecasting geomagnetic storms by exploiting machine learning (ML) approaches. By analyzing in-situ measurements of solar wind plasma and magnetic field data acquired from 2005 to 2019 at the Lagrangian point L1, we aim to predict one hour in advance a decrease of the SYM-H index below -50nT, which typically indicates geomagnetic disturbances. To such scope we employ Long Short Term Memory (LSTM) neural networks, best suited for the analysis of time series. Besides the L1 physical variables, the magnetic helicity and the energy carried by solar transients are considered as input features of the network architecture. Both the prediction scores and further investigation on the feature relevance show that such quantities have strong predictive capabilities. To provide numerical evidence of this fact we explore different state-of-the-art ML methods as Lasso regression, Recursive Feature Elimination (RFE) or Support Vector Machines (SVMs) as well as recently investigated Greedy feature selection techniques that allow us to perform a selection of smaller subsets of relevant features.

Author(s): Enrico Camporeale, Andong Hu

Queen Mary University of London & University of Colorado, Boulder; University of Colorado

Abstract: We present a novel model to forecast ambient solar wind a few days ahead. The method combines a physics-based propagation model with a deep learning feature extraction based on a Fourier Neural Operator.
The methodological novelty of the method resides in exploiting the automatic differentiation feature of modern ML packages in order to back-propagate the error through the numerical discretization of the solar wind propagation model. In this way, a neural network is trained to find the optimal solar wind boundary coronal condition that, once propagated to 1 AU match real-time observation recorded by the ACE satellite.
Along with the technical details of the method we showcase a comprehensive validation and benchmarks.

SWR6.3 Wed 6/11 09:00-10:15, room C2B – Sofia

Author(s): Francesco Pio Ramunno

University of Applied Sciences Northwestern Switzerland (FHNW)

Abstract: Investigating the solar magnetic field is crucial to understand the physical processes in the solar interior as well as their effects on the interplanetary environment. We introduce a novel method to predict the evolution of the solar line-of-sight (LoS) magnetogram using image-to-image translation with Denoising Diffusion Probabilistic Models (DDPMs). Our approach combines “computer science metrics” for image quality and “physics metrics” for physical accuracy to evaluate model performance. The results indicate that DDPMs are effective in maintaining the structural integrity, the dynamic range of solar magnetic fields, the magnetic flux and other physical features such as the size of the active regions, surpassing traditional persistence models, also in flaring situation. We aim to use deep learning not only for visualisation but as an integrative and interactive tool for telescopes, enhancing our understanding of unexpected physical events like solar flares. Future studies will aim to integrate more diverse solar data to refine the accuracy and applicability of our generative model. Visit this Github page for the code and this Hugging Space for the interactive tool.

 

Author(s): Christoph Schirninger, Robert Jarolim, Benoit Tremblay, Andrés Muñoz-Jaramillo, Brianna Isola, Rituparna Curt, Gautier Bardi de Furtou, Anne Spalding

Institute of Physics, University of Graz, Austria; High Altitude Observatory, Boulder, USA; Laboratory for Atmospheric and Space Physics, Boulder, USA; Southwest Research Institute, Boulder, USA; University of New Hampshire, Durham, USA; Stanford University, California, USA; NASA Jet Propulsion Laboratory, California, USA; Trillium Technologies, Inc

Abstract: Solar eruptive events pose significant risks to planets and its environment but also spacecraft systems and astronauts on human exploration missions. Therefore, the solar corona is permanently monitored in EUV where this layer appears highly structured. However, estimating the 3D  geometry is challenging due to the optically thin atmosphere. As part of the Frontier Development Lab 2022, we developed a method that provides novel 3D reconstructions of the solar EUV corona with neural radiance fields. The method applies radiative transfer and volume rendering techniques, to combine observations from multiple vantage points into a complete 3D representation using simulation data as well as data from SDO/AIA and STEREO/EUVI.
In this presentation, we showcase the outcomes of the FDL-X 2024 challenge, where we extend this approach to provide plasma diagnostics of the solar corona. We combine STEREO/EUVI, SDO/AIA and Solar Orbiter/EUI  observations, to provide a full 3D reconstruction of the Sun, which allows to estimate potential space weather risks, before the rotate into Earth-direction. Specifically, we will use this approach to predict spectral irradiance impacts on Mars, and use NASA’s MAVEN mission for verification. With this we improve our understanding of the plasma distribution in the solar corona, the total irradiance output, the dynamics of solar eruptions, as well as their potential impacts on Earth, Mars and the interplanetary space, which is essential for future manned space missions.

Author(s): Simone Chierichini, Greegoire Francisco, Jiajia Liu, Marianna Korsós, Robertus Erdélyi, Dario Del Moro

University of Sheffield; University of Rome ‘Tor Vergata’; University of Science and Technology of China; University of Sheffield; University of Sheffield; University of Rome ‘Tor Vergata’

Abstract: Space weather phenomena, particularly Coronal Mass Ejections (CMEs), have garnered significant scientific interest due to their potential to disrupt human activities on Earth. Machine learning has become popular for space weather predictions because it can extract patterns directly from data to produce accurate forecasts. Our study expands the CME Arrival Time Prediction Using Machine learning Algorithms (CAT-PUMA, Liu et al., 2018) with supervised learning techniques to improve predictions of CME arrival at Earth. We implemented and compared various supervised regression and classification models to predict CME transit times and classify CMEs as Earth-impacting or non-impacting. The study provides quantitative insights into forecasting Earth-impacting CMEs by utilising physical features and solar wind conditions at launch. Additionally, model interpretation techniques such as Shap Values were employed to identify and address the limitations of these machine learning models, as well as to gain insights into the feature space used to encode CME events. Finally, we explore how to leverage multimodal observations with deep learning to improve space weather forecasts and discuss how this approach can enhance CME transit time predictions.

Author(s): Maike Bauer, Justin Le Louëdec, Tanja Amerstorfer, Luke Barnard, Jackie Davies

Austrian Space Weather Office, GeoSphere Austria; Austrian Space Weather Office, GeoSphere Austria; Austrian Space Weather Office, GeoSphere Austria; University of Reading, Reading, UK; RAL Space, Rutherford Appleton Laboratory, Didcot, UK

Abstract: The capability to accurately predict the arrival of coronal mass ejections (CMEs) in real-time is crucial in mitigating the potential impact of severe space weather events. Detection and tracking of CMEs as they traverse the heliosphere is currently done mostly manually. Our objective is to use a convolutional neural network (CNN) for automatic detection and tracking of CMEs that can effectively utilize data from various heliospheric imaging (HI) instruments. The increasing number of active spacecraft throughout the heliosphere (including STEREO-A, Parker Solar Probe, and Solar Orbiter) equipped with HI instruments offers us an unprecedented opportunity to train our model using datasets covering a variety of viewpoints. The development and testing of such an algorithm holds particular significance at this time, with the launch of the PUNCH spacecraft in 2025 as well as the start of ESA’s Vigil mission planned for the early 2030s. PUNCH will be launched into a Sun-synchronous orbit and deliver high-resolution images of the space around the Sun from 5.9 out to 180 solar radii. Vigil will serve as a space weather monitor stationed at the L5 point, providing continuous, near real-time HI observations along the Sun-Earth line.
We present results from an automated CME detection and tracking algorithm trained on STEREO-HI observations. To improve temporal consistency during CME tracking, we use sequences of images to train our model. For model validation, we compare the time-elongation profiles it produces to those from the Solar Stormwatch project, a citizen science initiative that focused on tracking CMEs in STEREO-HI data. This comparison helps verify the accuracy and reliability of our model. Additionally, we provide an overview of potential future steps and challenges in the development of automated CME tracking systems.

Author(s): Simon Mackovjak, Adam Majirský, Silvia Kostárová, Samuel Amrich

Institute of Experimental Physics, Slovak Academy of Sciences; Faculty of Electrical Engineering and Informatics, Technical University of Košice; Institute of Experimental Physics, Slovak Academy of Sciences; Institute of Experimental Physics, Slovak Academy of Sciences

Abstract: The ESA mission Vigil will monitor space weather activity with the ambition to provide a timely warning of extreme solar events. Intending to enhance the reliability and timeliness of Vigil mission predictions, we have started a study that is focused on the most extreme events by using the data-driven approach. The ultimate goal of this study is to answer the following question by using Machine Learning (ML) techniques: Would it be possible to predict the occurrence of extreme space weather events, such as ‘Halloween storms’, if the Vigil mission was in operation at that time? During the talk, we will present the methodology for the identification of the most extreme space weather events in the last 30 years. The concept of Vigil-like instruments will be introduced to describe existing space-based instruments with capabilities similar to those planned for Vigil. The pipelines that might simplify access and processing of these data for the community will be presented, too. Finally, we will outline particular ML tasks that are in development within the ongoing study supported by ESA through the RPA program in Slovakia.

SWR6.4 Thu 7/11 09:00-10:15, room C2B – Sofia

Author(s): Ekaterina Dineva, George Miloshevich, Giovanni Lapenta, Jasmina Magdalenić Zhukov

KU Leuven, CmPA; KU Leuven, CmPA; KU Leuven, CmPA; KU Leuven, CmpA

Abstract: Contemporary solar physics deals with the increasing amount of high-dimensional data in historical data archives, as well as the constant influx of new observations, making it an excellent case for the application of artificial intelligence (AI) and, more specifically, machine learning (ML) algorithms. Synoptic full-disk observations with the Solar Dynamics Observatory (SDO) are one example, allowing us to follow the solar magnetic activity over more than one solar activity cycle and to study its local and global facets. The Space-weather HMI Active Region Patches (SHARP) vector magnetic field (VMF) maps and parameters, based on Helioseismic and Magnetic Imager (HMI) observations, have been developed to study magnetic evolution and flare triggering mechanisms. Empirical data such as SHARPs are a driving force for the development of flare prediction models. In this work, we aim to supply robust ML features based on the empirical SHARP dataset as a complementary input to a supervised flare forecasting model. Time series of SHARP VMF maps are used as input for Disentangled Variational Autoencoder (VAE), a type of Disentangled Representation Learning (DRL) algorithm, which facilitates the extraction of low-dimensional feature representation. The power of the VAE model lies in its ability to extract generalized information about nonlinear dynamical systems, in this case a solar active region, where each feature represents a particular aspect of the input data. The geometric features extracted by the VAE on its latent space are combined with usual SHARP parameters to make predictions of flare occurrences on a hold-out dataset. This method combines representation learning (unsupervised) and supervised methods. We compare this with purely supervised approaches common in the literature.

Author(s): Moritz Meyer zu Westram

University of Bern

Abstract: Solar flares and accompanying coronal mass ejections are drivers of intense space weather, which can have major impacts on e.g., satellite communication, navigation, and power-grid integrity. To this day, precise predictions of solar flare events remain challenging, due to the complexity of the underlying physical processes.
This study aims to improve solar flare forecasting through the application of survival analysis, a method traditionally used in fields like medicine and economics to model the timing of events and their related data features. In extension to previous studies, we aim to model not only the likelihood of a flare happening within the next few days but also its timing.
We demonstrate the time-to-event prediction capabilities of deep survival neural networks based on multivariate time series extracted from solar photospheric vector magnetograms in Spaceweather HMI Active Region Patch (SHARP) series.
Preliminary results indicate that deep survival analysis provides a promising new avenue for more precise event time predictions of solar flare outbursts. We found that including active regions that produce multiple flares in both training and validation sets, while keeping the flares themselves separated, yields highly accurate predictions with hour-level precision.

Author(s): Stefan Lotz

SANSA

Abstract: In this work we show how transfer learning can be applied to a prediction problem where forecasts of solar wind characteristics, such as those obtained from a Wang-Sheely-Arge (WSA) type model, are used to drive a deep learning based model of geomagnetic disturbance amplitudes.
Typically the prediction of ground based measures of geomagnetic disturbance relies on input from solar wind observations taken at the L1 point (SW-L1). However, these forecasts are of limited practical use for affected industry partners as the forecast lead time is typically only tens of minutes.
Forecasts of solar wind parameters near 1 AU (SW-FC) can be made utilising empirical and MHD based models. Typically these predictions are made shortly after CME eruption, thus yielding a lead time close to the estimated propagation time of the ejecta. GMD forecast lead times can be increased by driving models derived from SW-L1 data with the predicted solar wind parameters SW-FC.
In this work we utilise both SW-L1 and SW-FC data sets to develop a pair of models to predict GMD at multiple lead times. The source model is trained and tested on SW-L1 data to predict a GMD index based on magnetic field perturbations during moderate to extreme geomagnetic storms. The source model is then further trained and fine-tuned on a problem that has SW-FC data as input and the GMD index as output. This provides two models that can serve as a forecast-nowcast pair, driven by predicted and real-time solar wind data.

Author(s): Mario Cobos-Maestre

Universidad de Alcalá

Abstract: Machine Learning has gained favor as a tool for space weather event forecasting over the last few years.  Deep neural networks in particularly have been typically leveraged to consume large amounts of data to automatically estimate the behavior of relevant measurements, such as radio emission flux or impact on geomagnetic indices.
In the context of solar wind forecasting, efforts have been primarily focused on forecasting the bulk speed of solar wind streams. Although steps have been taken in improving the accuracy of forecasts, a common concern from an operational point of view  has been determining the reliability of any given prediction.
In this work, we explore the predictive ability obtained by using a variational neural network with a normal distribution function output as a forecasting tool. We find that the resulting model significantly reduces error compared to state-of-the-art neural networks in the published literature, achieving an average RMSE of 31.30 km s^−1 in the test set over several realizations.
We also leverage the properties of the Bayesian inference process underpining the resulting model to estimate the degree and sources of uncertainty for any given forecast. We provide a theoretical 95% confidence interval for each prediction computed from the standard deviation of the forecasted normal distribution.  For a lead time of five days, we find that the empirical accuracy of this interval is of 94.72% in the test set.

Author(s): Daniel Collin, Yuri Shprits, Stefan Hofmeister, Stefano Bianco, Guillermo Gallego

GFZ German Research Centre for Geosciences; Technical University of Berlin; GFZ German Research Centre for Geosciences; University of Potsdam; University of California Los Angeles; Columbia University; GFZ German Research Centre for Geosciences; Technical University of Berlin; Einstein Center Digital Future

Abstract: The solar wind is driven by large structures on the solar surface like coronal holes and active regions, which can be identified in extreme ultra-violet (EUV) solar images several days before they become geoeffective. In this work, we propose to use a distributional regression algorithm to forecast the solar wind speed at the Lagrange 1 point from solar images. Instead of predicting a single value, this method models the entire conditional distribution as a function of input features. It allows specifying the probability of the solar wind speed exceeding certain thresholds, which is especially useful for extreme event predictions like coronal mass ejections and high-speed solar wind streams. First, we construct a feature extractor that encodes each multispectral image into a low-dimensional representation. These representations are aggregated over a time span of multiple days to capture information about the evolution of the features and are paired with other physically relevant parameters (e.g., past solar wind properties and solar cycle information). The collected features are passed to a generalized additive model for location, scale and shape (GAMLSS) to predict the hourly solar wind speed multiple days ahead. We show that through this approach we can capture the distribution of the solar wind speed, including the heavy tails, and significantly improve the accuracy of extreme event prediction. We compare our method to other algorithms in terms of several metrics over 14 years of data and show the advantages of the availability of a probability distribution for each prediction.

Posters

Posters I  Display Tue 5/11 – Wed 6/11, room C1A – Aeminium

Authors in attendance: Tue 5/11 10:15–11:30, 15:15-16:15; Wed 6/11 10:15–11:30

Author(s): Andy Smith, Jonathan Rae, Colin Forsyth, John Coxon, Maria Walach, Christian Lao, Shaun Bloomfield, Sachin Reddy, Mike Coughlan, Amy Keesee, Sarah Bentley

Northumbria University; Northumbria University; UCL; Northumbria University; Lancaster University; UCL; Northumbria University; UCL; UNH; UNH; Northumbria University

Abstract: Geomagnetically Induced Currents (GICs) are a severe space weather hazard, driven through coupling between the solar wind and magnetosphere.  GICs are rarely measured directly, instead the ground magnetic field variability is often used as a proxy.  Recently space weather models have been developed to forecast whether the magnetic field variability (R) will exceed specific, extreme thresholds.  We test an example machine learning-based model developed for the northern United Kingdom.  We evaluate its performance (discriminative skill and calibration) as a function of magnetospheric state, solar wind input and magnetic local time.
We find that the model’s performance is highest during active conditions, for example geomagnetic storms, and lowest during isolated substorms and “quiet” intervals, despite these conditions dominating the training dataset.  Correspondingly, the performance is high when the solar wind conditions are elevated (i.e. high velocity, large total magnetic field strength, and the interplanetary magnetic field oriented southward), and at a minimum when the north-south component of the magnetic field is highly variable or around zero.  Regarding magnetic local time, performance is highest within the dusk and night sectors, and lowest during the day.  The model appears to capture multiple modes of magnetospheric activity, including substorms and viscous interactions, but poorly predicts impulsive phenomena (i.e. storm sudden commencements) and longer timescale coupling processes.
We discuss the implications of these results for future forecasting efforts.

Author(s): Vincenzo Timmel, Prof. Dr. André Csillaghy, Christian Monstein

FHNW; FHNW; Istituto ricerche solari Aldo e Cele Daccò (IRSOL), Università della Svizzera italiana, Locarno, Switzerland.

Abstract: Monitoring solar activity at radio frequencies is recognised as a key component of space weather operations. Solar Radio Busts (SRBs) can be used as an early warning system for solar storms.
The e-Callisto network provides a global radio monitoring capability with observing stations conveniently located around the world, covering the range of 45 MHz to 870 MHz, or any other range by inserting a frequency converter.
The observations are radio spectrograms. They provide a consistent basis for burst detection, but this is a challenging task due to the diversity of antenna characteristics, radio frequency interference and low signal-to-noise ratio associated with each observing station. Therefore, we are investigating machine learning algorithms to increase robustness.
In the study presented here, we build a training data set consisting of 15-min spectrograms with binary labels: True if the spectrogram contains bursts and False if it does not. The labels are taken from a burst list manually compiled by an expert. SRBs make up 9% of the dataset, reflecting the relative rarity of such events. To prevent the model from focusing on a particular position within the spectrogram, the start time of the spectrogram is randomised by several minutes from the actual burst occurrence. We further improve the robustness of the model to out-of-distribution samples using data augmentation techniques borrowed from natural language processing, in particular time-warping and SpecAugment, which distort spectrograms in time and mask frequencies, respectively. The model itself is based on a simple ResNet architecture to capture data patterns. Despite some incorrect labels in the training and validation sets, our curated test set shows that the model achieves an F1 score of 92.8%, a precision of 91.8% and a recall of 92%, with a class balance of 1:8. This demonstrates its effectiveness in distinguishing burst from non-burst events.
In addition, the model performs well on stations that do not appear in the training data, achieving a precision of 91% and a recall of 89.3%. 

Author(s): Roksoon Kim, Jongyeob Park, Woong Jeon

KASI; KASI; KHU

Abstract: Identifying interplanetary coronal mass ejections (ICMEs) in solar wind observations is crucial for scientific objectives, such as understanding the evolution and propagation of CMEs and the embedded magnetic flux rope structures. It is also essential for practical purposes, such as comprehending their impact on geomagnetic storms. However, the observation of ICMEs in solar wind data is highly diverse, making it challenging to promptly detect the onset of ICME structures, which signifies their arrival at Earth. In this study, we aim to develop an ICME detection algorithm using AI techniques, utilizing ACE solar wind observations and the Richardson and Cane ICME list from 1998 to 2022. Through this effort, we intend to lay the foundation for a real-time ICME automatic detection algorithm for future space weather forecasting. We believe the ICME detection lists generated in this manner will prove invaluable for the physical study of CME-ICME interactions.

Author(s): Hannah Theresa Rüdisser, Gautier Nguyen, Christian Möstl, Justin Le Louëdec, Eva Weiler, Emma E. Davies, Ute V. Amerstorfer

Austrian Space Weather Office, GeoSphere Austria, Graz, Austria; ONERA/DPHY, Université de Toulouse, Toulouse, France; Austrian Space Weather Office, GeoSphere Austria, Graz, Austria; Austrian Space Weather Office, GeoSphere Austria, Graz, Austria; Austrian Space Weather Office, GeoSphere Austria, Graz, Austria; Austrian Space Weather Office, GeoSphere Austria, Graz, Austria; Austrian Space Weather Office, GeoSphere Austria, Graz, Austria

Abstract: Interplanetary coronal mass ejections (ICMEs) and stream interaction regions (SIRs) are primary drivers of space weather disturbances. Various approaches have been used in the past to automatically detect events in time series data from solar wind in situ observations. However, accurate and fast detection remains challenging given the large volumes of data from different instruments, as well as the lack of community-agreed definitions. In recent years, deep-learning based methods emerged, dedicated to the automatic detection of those events from archived L1 in situ data. We now aim to extend their application to a real-time, multi-class detection scenario to evaluate its potential as an early warning system. Real-time data stream detection is generally more complicated, as one has to deal with low quality data and the fact that only parts of the structure have arrived by the time a detection would be needed. To address these challenges, we combine recent efforts with physical models, such as CME arrival time models, and additional data sources including coronal CME catalogs or HI catalogs.

Author(s): Aishwarya Balivada, Fech Scen Khoo, Dominika Malinowska, Hannah Theresa Rüdisser, Filip Svoboda, Daniele de Martini, Giacomo Acciarini, Martin Sanner, Carlos Urbina Ortega, Manuel Lacal, Giuseppe Mandorlo

Department of Physics and Astronomy, Purdue University; Institute of Physics, University of Oldenburg; Department of Architecture and Civil Engineering, University of Bath; Department of Geoscience and Engineering, Delft University of Technology; Austrian Space Weather Office, GeoSphere Austria; Institute of Physics, University of Graz; CaMLSys, University of Cambridge; Oxford Robotics Institute, University of Oxford; Advanced Concepts Team, European Space Agency, European Space Research and Technology Centre (ESTEC); Surrey Space Centre, University of Surrey; School of Science and Engineering, University of Dundee; TEC-ED, European Space Agency; TEC-ED, European Space Agency; TEC-ED, European Space Agency

Abstract: We present the development of a machine-learning-based predictive system for solar eruptions, providing early warnings through satellite data. The ESA VIGIL mission and its successors, due to launch to Lagrange Point 5 within the next decade, will offer an early view of the Sun, with a five-day advanced warning. This is vital for mitigating potential damages and disruptions caused by space weather.
This work stems from an eight-week research sprint conducted as part of the Frontier Development Lab Europe 2024. During this period, we analysed diverse data sources, such as magnetic fields, EUV, and coronagraphic images, in conjunction with solar wind data. This analysis aims to suggest relevant instruments for VIGIL 2.0 to advance solar flare prediction and extend the lead time before solar eruptions impact Earth.
Our predictive system integrates data from multiple satellites, including VIGIL surrogates such as the ESA’s Solar Orbiter, Proba-2, NASA Stereo, and NASA ACE, focusing on studying solar eruptions. This multi-modal, multi-view approach aims to enhance our understanding of solar processes and, thus, solar eruptions. By doing so, we analyse the effectiveness of different sensor modalities in pursuit of an early warning of space weather, potentially leading to relevant new experiments to consider.

Author(s): Richard Boynton

University of Sheffield

Abstract: Space weather forecasting models for the radiation belts are deduced using the system identification method based on Nonlinear AutoRegressive Moving Average eXogenous (NARMAX) models.  NARMAX methods, akin to machine learning methods, detect a model from input-output data, which can then be implemented to provide a forecast. In the SWIMMR N1 project, Rad-Sat, which was devoted to forecasting satellite risk, NARMAX was applied to deduce models for the electron fluxes at geostationary orbit between 50 keV and 4 MeV, the Kp Index, and electron counts between L = 1.5-8 RE for >30 keV, >100 keV and >300 keV, using solar wind parameters as the inputs. The forecasts of these models are compared to the spacecraft and geomagnetic index measurements using a number of performance metrics. These models were then incorporated into a real time forecasting system for the Met Office.

Author(s): Qusai Al Shidi

West Virginia University

Abstract: Physics based models are often used to predict activity in the ring current during geomagnetic storms. These physics-based models are computationally expensive. We can use machine/deep learning methods to develop an emulator, that we call a reduced order probabilistic emulator (ROPE), in the reduced state space to overcome this limitation. The two primary steps of this workflow are dimensionality reduction to obtain the latent or reduced state space, and time-series modeling in the reduced state space. Previously, we have used Principal Component Analysis (PCA), a linear method, to compress (step 1) the omni-directional flux output of the Ring current Atmosphere Interactions Model with Self Consistent magnetic field (RAM-SCB) truncated to 20 principal components. In this work, we extend upon this work by employing a nonlinear autoencoder (AE), providing superior performance with a latent dimension of 10.  We train, validate and test on the same 30 simulation periods of geomagnetic storms used in the previous PCA study. We show the improved performance of using a non-linear technique like an AE in comparison with the previously used linear method, PCA, by comparing the errors. Combined with the time-series modeling step as previously demonstrated, this results in a highly capable ROPE that can provide forecasts with robust and reliable uncertainty quantification using ensemble methods.

Author(s): Piyush Mehta, Richard Licata

West Virginia University; West Virginia University

Abstract: Physics-based models are ubiquitous across space weather. One of the primary challenges limiting practical implementation of physics-based models in science and operations is its computational cost. Reduced order emulators can alleviate this limitation and retain majority of the fidelity of the underlying physical model. We have demonstrated the application of multiple AI/ML algorithms for the development of Reduced Order Probabilistic Emulators (ROPEs) across geospace domains – thermosphere, ring current, and ground magnetic perturbations. For the thermosphere, developing a ROPE is the missing step towards operationalizing physics-based models. We present recent advances in this area supported by the US Office of Space Commerce. The two primary steps of this workflow are 1) dimensionality reduction to obtain the latent or reduced state space, and time-series modeling in the reduced state space. Previously, we have used Principal Component Analysis (PCA), a linear method for dimensionality reduction (step 1).  Additionally, the model was limited to an altitude of 450 km. In this work, we alleviate both these limitations by extending the model to 1000 km altitude and using the nonlinear autoencoder to significantly improve performance during active conditions. This results in a highly capable ROPE that can provide forecasts with robust and reliable uncertainty quantification using ensemble methods.

Author(s): Carl Shneider, Andreas M. Hein, Vasily L. Petrov

University of Luxembourg, SpaSys Group; University of Luxembourg, SpaSys Group; Mission Space

Abstract: Satellites in LEO orbit are exposed to high energy radiation, more so on account of geomagnetic storms which exacerbate harmful space environmental effects. For example, high energy radiation belt photons cause ionization effects and displacement damage effects – harmful for opto-electronic technologies, while high energy electrons contribute to deep dielectric charging and discharging resulting in performance degradation of electronic equipment and computer bit flips. In this study we employ information theoretic measures in parallel with machine learning for time series data techniques to make a 2-4 day ahead forecast of the energy fluxes of high energy protons and electrons averaged over L-shells L2 – L4. First, we recast storm times in terms of the new composite magnetospheric activity index: the whole-Earth index E(1) proposed by Borovsky and Lao (2023) in place of the traditionally used solitary Dst or Kp indices. Secondly, we access mutual information between combinations of upstream solar wind variables and magnetospheric indices to arrive at a set of parameters that can separate storm times from quiet times. Finally, with these assessments, we compare with results output by several architectures of neural networks.

Author(s): Ehsan Tavabi, Rayhaneh Sadeghi

PNU; PNU

Abstract: Bright Points (BPs) are a significant aspect of solar observations. This study aimed to investigate the characteristics of BPs in different regions of the Sun, with a specific focus on their oscillatory behavior. Recent advancements in machine learning and image processing have revolutionized the field of solar physics (Sadeghi & Tavabi 2022a,b; Tavabi 2018). A machine learning model was developed to identify and analyze BPs in solar images. The model was trained using over 2,000 slit jaw images (SJIs) from IRIS and achieved 78% accuracy in BP identification. Oscillatory behavior of the identified BPs was further analyzed using wavelet and Fourier analysis techniques. The image processing pipeline involved steps such as image enhancement, segmentation, feature extraction, and tracking. Accurate BP identification was achieved using a convolutional neural network (CNN) that learned patterns and features to predict BP brightness and coordinates. Data augmentation techniques were employed to mitigate overfitting. This pipeline provided a flexible framework for customized analysis based on the dataset and research goals. Overall, the study demonstrated an effective approach to analyzing BPs in solar images by integrating machine learning and oscillation analysis techniques. The researchers observed both differences and similarities in the properties of oscillated and non-oscillated BPs across various solar regions, including the quiet Sun (QS), active regions (ARs), and coronal holes (CH). The damping per period, known as the Q-factor, and the maximum Doppler velocity of BPs varied depending on the region under investigation. In the quiet Sun, internetwork BPs exhibited shorter damping times (120 seconds) and higher maximum Doppler velocities (47 km/s) compared to network BPs. Network BPs in the quiet Sun had damping times of 216 seconds and maximum Doppler velocities of 37 km/s. In active regions, internetwork BPs generally had longer damped times (around 220 seconds) and wider ranges of maximum Doppler velocities (10 to 140 km/s) compared to network BPs, which had damping times of 130 seconds and maximum Doppler velocities ranging from 10 to 85 km/s. In coronal holes, both types of BPs displayed similar damping times of 120 seconds, but internetwork BPs tended to have higher maximum Doppler velocities (100 km/s) compared to network BPs (85 km/s). The study also highlighted that most network BPs in active regions exhibited overdamping behavior, indicating a stronger damping effect. In the quiet Sun, internetwork BPs showed overdamping behavior, while oscillated network BPs displayed critical damping behavior. However, it is essential to note that the specific local plasma conditions and magnetic environment may influence the physical mechanisms responsible for BP damping. By investigating these properties, the study provided insights into the characteristics and behaviors of BPs in different solar regions, emphasizing their complex nature and the need to consider region-specific conditions when studying their oscillatory behavior and damping mechanisms. This research serves as a foundation for unraveling the complexities of BPs and their role in solar activity, which has implications for space weather forecasting and our comprehension of the Sun-Earth relationship (Sadeghi & Tavabi 2024a,b; Tavabi & Sadeghi 2024a,b;).

Author(s): Rayhaneh Sadeghi, Ehsan Tavabi

PNU; PNU

Abstract: Solar bright points (BPs) are intense emission regions observed in the chromosphere and transition region of the Sun. Understanding their dynamics and energy dissipation processes is crucial for unraveling the complex behavior of the Sun’s outer atmosphere. In this study, we utilized spectral analysis and deep learning techniques to examine the damping characteristics of Doppler velocity oscillations in BPs. Using data from the Interface Region Imaging Spectrograph (IRIS), we analyzed the Doppler shift in the solar spectrum, specifically focusing on periodic oscillations within BPs in the chromosphere and transition region. We investigated the development of longitudinal oscillations with damping and observed the damping of red and blue Doppler shifts at different time intervals. Additionally, we employed deep learning techniques to explore the statistical properties of damping in network and internetwork BPs within active regions and coronal hole areas (Sadeghi & Tavabi 2022a; Sadeghi & Tavabi 2022b; Tavabi et al. 2022; Tavabi 2018). Our results revealed significant variations in damping rates across different regions. The highest damping was observed in the network BPs within coronal holes, indicating rapid energy dissipation. Internetwork regions exhibited shorter decay times and half-lives compared to network regions, suggesting a higher damping rate. Moreover, coronal hole regions displayed shorter decay times and half-lives compared to active regions, which can be attributed to lower density and weaker magnetic fields in coronal hole areas. Time-Doppler slices derived from IRIS Si IV (1400 Å) revealed heightened acceleration values in network BPs compared to other locations, indicating their significant role in energy transport and heating processes in the solar atmosphere. Our findings provide insights into the damping behavior of BPs in different solar regions. The observed underdamped nature of the oscillations suggests that BPs possess sufficient energy to drive the fast solar wind and contribute to heating the quiet corona. The presence of rapid damping in internetwork regions and coronal hole areas suggests the influence of small-scale magnetic fields and lower plasma densities in these regions. Additionally, we discovered a potential association between network BPs and the footpoints or cross sections of spicules, prominent features of the chromosphere. This finding supports the idea of a connection between BPs and spicular activity, although further research is needed to fully comprehend this relationship (Sadeghi & Tavabi 2024; Tavabi & Sadeghi 2024). Our study highlights the dynamic nature of solar BPs and provides valuable insights into their damping characteristics and energy dissipation processes. The variations in damping rates across different solar regions contribute to our understanding of the complex dynamics of the Sun’s outer atmosphere. These findings enhance our knowledge of energy transport, heating mechanisms, and the overall behavior of the solar chromosphere and transition region. Future research in this area, utilizing advanced observational and computational techniques, along with deep learning approaches, will further advance our understanding of the Sun’s dynamic behavior and its impact on space weather phenomena. Our study contributes to the ongoing efforts to unravel the mysteries of the Sun and provides valuable insights for future investigations in solar physics.

Author(s): Guram Kervalishvili, Ingo Michaelis, Jan Rauberg, Monika Korte, Jürgen Matzka

GFZ German Research Centre for Geosciences, Potsdam, Germany; GFZ German Research Centre for Geosciences, Potsdam, Germany; GFZ German Research Centre for Geosciences, Potsdam, Germany; GFZ German Research Centre for Geosciences, Potsdam, Germany; GFZ German Research Centre for Geosciences, Potsdam, Germany

Abstract: The Kp and ap indices, derived and distributed by the GFZ German Research Centre for Geosciences, are widely used to measure geomagnetic activity in various applications. These applications include assessing geomagnetic disturbance levels, creating models for the near-Earth space environment, and analyzing data. The Kp index has been available since 1932 and is particularly important for studying long-term space climate trends. In addition, GFZ has introduced a new set of geomagnetic indices called Hpo (“H” stands for half-hourly or hourly measurements, “p” represents planetary, and “o” indicates open-ended). These indices have been calculated back to 1985 and offer a detailed view of planetary geomagnetic activity similar to the Kp index but without an upper limit. The Hpo index family consists of the half-hourly Hp30 and ap30 indices, as well as the hourly Hp60 and ap60 indices.
Predicting geomagnetic activity is crucial for mitigating the risks linked with space weather and its significant effects on modern technological infrastructure. Enhanced geomagnetic activity has the potential to heighten satellite orbital drag, cause radiation harm to spacecraft, disrupt satellite communications and navigation systems, impact power grids, and accelerate pipeline corrosion.
We have developed a new prediction model for the Kp, ap, Hp60, ap60, Hp30 and ap30 indices utilizing machine learning methods. This model uses solar wind parameters, sunspot number, interplanetary magnetic field measurements, and past geomagnetic indices. Despite being simple, the new model delivers reliable forecasts based on past geomagnetic and solar wind data. We are currently validating real-time forecast models and evaluating various solar wind forecast models that account for high-speed solar wind streams as well as coronal mass ejections. Our initial findings suggest that the proposed model effectively predicts the Kp, ap, Hp60, ap60, Hp30, and ap30 indices for both moderate and severe events, capturing the dynamic nature of these indices.

Author(s): Jihyeon Son, Yong-Jae Moon

Department of Astronomy and Space Science, Kyung Hee University, Yongin, 17104, Republic of Korea; Department of Astronomy and Space Sciece, Kyung Hee University, Yongin, 17104, Republic of Korea / School of Space Research, Kyung Hee University, Yongin, 17104, Republic of Korea

Abstract: We have developed two deep learning models to forecast two space weather components: 3-day solar wind speeds and 6-hour interplanetary magnetic field (IMF) Bz components. Firstly, the solar wind speed prediction model uses the last five days of SDO/AIA 19.3 and 21.1 nm images, along with solar wind speeds as input data. It consists of two networks: a convolutional layer-based network for images and a dense layer-based network for solar wind speeds. Our model successfully predicts solar wind speeds for the next 3 days, with a root mean square error (RMSE) ranging from 37.4 km/s (6-hour prediction) to 68.2 km/s (72-hour prediction). These results are much better than those of previous studies. The model can accurately predict sudden increases in solar wind speeds caused by equatorial coronal holes. Secondly, the Bz prediction model is a bidirectional long short-term memory (BiLSTM) based model using solar wind data (V, N, T) and IMF (Bt, Bx, By, Bz) in OMNI from 2000 to 2022 as input data. We use the preceding 12 hours of data as input and the next 6 hours of data as target data. We consider Bz values below the negative standard deviation (about -3 nT) for at least 6 hours. We apply 12-fold cross-validation to our model, using 8 months for training sets and 4 months for test sets. Consequently, a total of 12 models are trained, and they show an averaged RMSE ranging from 1.75 nT (30-minute prediction) to 2.55 nT (6-hour prediction). Our model can capture both declining and increasing phases of Bz. Although this study presents preliminary results in Bz prediction, we find a sufficient possibility for predicting Bz under specific conditions. We plan to develop deep learning models for other space weather components, such as solar wind density or geomagnetic indices.

Author(s): Mahmoud Mohamed

Cairo university

Abstract: The estimation of the solar magnetic fields on the far side of the Sun is crucial for advancing our understanding of solar dynamics and space weather forecasting. In this study, we present a Deep Learning (DL) approach for estimating the solar far-side magnetic field. The model was trained on full disk synchronized images from the Solar Terrestrial Relations Observatory/Extreme Ultraviolet Imager (STEREO/EUVI), the Solar Dynamics Observatory/Atmospheric Imaging Assembly (SDO/AIA) at 304Å, and in conjunction with data from the SDO/Helioseismic and Magnetic Imager (HMI).The SDO/AIA EUV observations were homogenized with respect to the STEREO/EUVI ones. The DL model is trained on the homogenized SDO/AIA and SDO/HMI observations to estimate the magnetic field using only the EUV data. That model is used with the STEREO/EUVI observations to estimate the far-side magnetic field. Preliminary results demonstrate the feasibility and effectiveness of the proposed approach in estimating the solar far-side magnetic field, providing valuable insights for space weather prediction applications.

Author(s): Kendra R. Gilmore, Sarah N. Bentley, Andy W. Smith

Northumbria University, Newcastle-Upon-Tyne, UK; Northumbria University, Newcastle-Upon-Tyne, UK; Northumbria University, Newcastle-Upon-Tyne, UK

Abstract: The aurora has always been a phenomenon that has fascinated humans throughout history and still does. However, with our increased dependency on our technological infrastructure, which is susceptible to external electromagnetic changes related to the location of the aurora, we need to predict the location of the aurora accurately.
Current auroral prediction models rely on our understanding of the interaction between the magnetosphere and the solar wind or geomagnetic indices. Both approaches are well established but have limitations with respect to forecasting (geomagnetic indices-based model due to a need of index forecast first) or because of the underlying assumptions driving the model (due to a simplification of the complex magnetosphere-solar wind interaction). Machine learning (ML) provides an alternative approach to this question with its strongly data-driven nature and the possibility of capturing non-linear behaviour.
We train a Long-Short Term Memory (LSTM), a type of ML algorithm well suited to time series, with ground magnetic field and solar wind data as input and auroral boundaries derived from IMAGE satellite UV images as labels. We also investigate to what extent periodic behaviour of input variables can be captured by an LSTM from purely temporal data using synthetic data and if including extra information on intrinsic cycles can improve performance. Incorporating those results into the auroral LSTM model, we develop an initial ML-based model for auroral prediction.

Author(s): Nuno Costa, Filipa S. Barros, João J. G. Lima, Rui F. Pinto, André Restivo

LIACC / Faculdade de Engenharia da Universidade do Porto, Portugal; LIACC / Faculdade de Engenharia da Universidade do Porto; Instituto de Astrofísica e Ciências do Espaço, CAUP; Institut de Recherche and Astrophysique et Planétologie, OMP/CNRS, CNES, University of Toulouse;; Instituto de Astrofísica e Ciências do Espaço, CAUP; Departamento de Física e Astronomia, Faculdade de Ciências da Universidade do Porto,; Institut de Recherche and Astrophysique et Planétologie, OMP/CNRS, CNES, University of Toulouse; LIACC / Faculdade de Engenharia da Universidade do Porto

Abstract: Space weather encompasses the dynamic conditions within the solar system, focusing on how the solar wind – a flow of charged particles from the Sun – affects Earth’s magnetic field and atmosphere. Reliable predictions of space weather are essential to protect satellites, communication networks, power grids, and the safety of astronauts. Current models like MULTI-VP are computationally demanding. We suggest using Physics-Informed Neural Architectures, such as Neural Networks and Neural Operators (PiNNs and PiNOs), which provide a quicker and still precise alternative that adheres to physical laws. These models integrate physics with data-driven methods to offer fast and accurate forecasts. Our research indicates that both PiNNs and PiNOs can decrease computation times and achieve forecasting results comparable to those of MULTI-VP, thus providing a more efficient and reliable method for predicting solar wind effects.

Author(s): Alberto Regadío, J. Ignacio G. Tejedor, Oscar García-Población, Juan José Blanco

Universidad de Alcalá; Universidad de Alcalá; Universidad de Alcalá; Universidad de Alcalá

Abstract: The detection of neutron produced by cosmic ray interaction with atmospheric nuclei plays a crucial role in Space Weather observation. In the development of electronics for neutron monitors (NMs) such as the Castilla-La Mancha NM (CaLMa), the size of the gas tubes, their location, and the low rate of events in a counter tube pose significant challenges. To address these issues, we propose a novel solution using artificial neural networks, specifically Generative Adversarial Networks (GANs), to generate simulated pulses that closely mimic the characteristics of real detector pulses. The primary objective of this work is to develop a GAN-based model capable of producing synthetic pulses indistinguishable from those generated by neutron monitors, thereby facilitating the testing and development of associated electronics without the need for actual detectors or preamplifiers.
The architecture of the GAN, including the generator and discriminator networks, is designed to capture the intricate features of real pulses from the CaLMa NM but can be easily adapted to other facilities or detectors. The training methodology involves using real pulse data to ensure that the generated pulses exhibit the same shape and statistical properties as the real ones. Once trained, the GAN model is deployed on a Xilinx System-On-Chip (SoC), specifically a Xilinx/AMD Pynq board. This hardware integration allows the synthetic pulse generator to be used as a direct replacement for the actual neutron detector and its associated preamplifier during electronics testing.
By using the Pynq board instead of the neutron tubes, we provide a compact, portable, and efficient solution for simulating detector pulses, which is particularly advantageous during the development and debugging phases of the detector electronics. Our results demonstrate that the GAN-generated pulses maintain the same morphological characteristics and match the data distributions in the original pulse-height histogram data, ensuring high fidelity to the real pulses.

Author(s): Marta Guimaraes, Maria Almeida, Tiago Baptista, Rob Arthur, Chiara Manfletti

Neuraspace; Neuraspace; Neuraspace; Neuraspace; Neuraspace

Abstract: Thermospheric density plays a critical role in precise orbit injection, orbit determination and propagation, collision avoidance, re-entry and re-entry estimation, and satellite lifetime and lifetime assessments. However, due to its stochastic nature, it is one of the main sources of uncertainty when estimating the satellites’ position and velocity in low-Earth orbit. Due to approximations, assumptions, and limited temporal resolution, the current empirical models used to estimate this variable do not provide sufficiently accurate estimates. Besides, such models are known to fail during extreme solar activity, e.g., during geomagnetic storms. Some researchers have resorted to machine learning techniques to avoid considering physics-based assumptions and leverage historical patterns to improve existing solutions. In this paper, we expand the work available in Karman, an open-source package presenting state-of-the-art solutions for thermospheric density estimation. More specifically, we propose an extended version of the machine learning model architecture aiming not only at improving its predictive power but also at providing confidence intervals associated with the predictions. Such intervals can aid in downstream tasks, e.g., orbit determination and propagation procedures, by providing an additional measure of uncertainty that should improve covariance realism. To better align with real-life satellite operational contexts, we propose changes to the training and evaluation setup as presented in Karman. To that end, we explore the use of different train/validation/test dataset splits in order to better simulate the use of such a model in production, ensuring the test dataset only contains future data in relation to the training dataset. Additionally, we test the generalisation capacity of the model with previously unseen density data from the GRACE-FO mission. The results show the potential of our approach to predict the thermospheric density and the importance of associating confidence intervals with the predictions. This work contributes to mitigating the overall risks posed by the limitations of the current thermosphere density estimation models and to the ongoing efforts to improve space traffic management and coordination. As we advance through periods of heightened solar activity, these models can prove even more relevant.

Author(s): Edoardo Legnaro, Sabrina Guastavino, Anna Maria Massone, Michele Piana, Shane Maloney, Paul Wright

Università degli Studi di Genova; Università degli Studi di Genova; Università degli Studi di Genova; Università degli Studi di Genova; Dublin Institute for Advanced Studies; Dublin Institute for Advanced Studies

Abstract: A solar active region disrupts the Sun–Earth space environment, often leading to severe space weather events. Sunspots are the primary markers of these active regions, and specific sunspot types are associated with space weather events such as solar flares and coronal mass ejections. Therefore, the automatic classification of sunspot groups is beneficial for accurately and promptly predicting solar activity.
A significant challenge in this problem is the inherent imbalance in the dataset of solar active regions. The distribution of sunspot types is highly skewed, with more complex sunspot structures being much rarer than simpler ones. This imbalance can lead to biased models that underperform in identifying the less common, yet potentially more critical, sunspot types associated with severe space weather events.
In this talk, we will present our results in applying deep learning techniques to the classification of sunspot magnetogram cutouts based on the Mount Wilson classification scheme. We will review the current state-of-the-art architectures in image classification, from Convolutional Neural Networks to Vision Transformers, and present the best network architectures for this task as well as our training pipeline. In addition, we will discuss the most effective techniques for dealing with the strong dataset imbalance, including data augmentation and appropriate choices of loss functions. This work has been supported by the Active Region Classification and Flare Forecasting (ARCAFF) project.

Author(s): Daniel Conde Villatoro, Florencia Castillo, Carlos Escobar Ibáñez, Carmen García García, Jose Enrique García Navarro, Verónica Sanz González, Bryan Zaldívar Montero, Juan José Curto Subirats, Santiago Marsal Vinadé, Joan Miquel Torta Margalef

Instituto de Física Corpuscular (IFIC), CSIC -Universitat de València; Laboratoire d’Annecy de Physique des Particules (LAPP), CNRS/IN2P3; Instituto de Física Corpuscular (IFIC), CSIC -Universitat de València; Instituto de Física Corpuscular (IFIC), CSIC -Universitat de València; Instituto de Física Corpuscular (IFIC), CSIC -Universitat de València; Instituto de Física Corpuscular (IFIC), CSIC -Universitat de València; Instituto de Física Corpuscular (IFIC), CSIC -Universitat de València; Observatori de l’Ebre (OE), University Ramon Llull – CSIC; Observatori de l’Ebre (OE), University Ramon Llull – CSIC; Observatori de l’Ebre (OE), University Ramon Llull – CSIC

Abstract: In recent decades, the reliance of modern society on various technological infrastructures has significantly increased, rendering them vulnerable to the effects of space weather. These effects are primarily driven by solar activities and associated events, such as coronal mass ejections and solar flares. Among other things, these phenomena produce geomagnetic storms that can disrupt Earth’s magnetosphere and ionosphere. These disturbances lead to geomagnetically induced currents (GICs) that can flow through technological systems, causing perturbations, interruptions, and even long-term damage to critical infrastructures such as power grids, railways, and buried gas pipelines, with dramatic social, economic, and political consequences.
This project aims to establish a robust early warning system to forecast the impact of space weather on critical infrastructures in Spain, enabling informed decisions and actions, and thereby enhancing preparedness and mitigation strategies. In order to successfully develop such a trustworthy early warning system and ultimately a vulnerability map, the intermediate and necessary goal of this project is to develop a highly reliable predictive model of the impact of future solar storms on the ground in Spain. This is a challenging task, given the highly non-linear dependencies of the magnetosphere’s response to space weather. Our model is being developed in two distinct stages. First, we are using real-time data from the solar wind space probe ACE (located at the L1 point in space) and data from the Observatori de l’Ebre to develop a deep-learning-based model. This model takes past conditions into account to predict the geomagnetic activity index SYM-H, and later the variation of the geomagnetic field on the Earth’s surface at this observatory. We have developed a robust model based on long short-term memory (LSTM) networks, which is currently being generalized and further improved. We are also boosting its performance and scalability by using state-of-the-art deep-learning algorithms such as Transformers. Second, in collaboration with our colleagues from the Observatori de l’Ebre, we are beginning to feed these local predictions of the magnetic field’s time variation into a physical model (the Electrical Resistivity Model for the Iberian Lithosphere (ERMIL) developed by our colleagues at the University of Barcelona) of the 3D Earth’s geoelectrical structure to generate the geoelectrical fields that drive the GICs. This dual approach will ensure that our model not only forecasts space weather events but also accurately assesses their potential impacts on the ground.
In conclusion, this project represents a significant step forward in developing autonomous, robust, and highly reliable space weather prediction systems in Europe. By leveraging advanced deep-learning methods, coupled with real-time data integration and a robust physical model, we aim to mitigate the risks posed by space weather events and GICs, ensuring the resilience and stability of critical infrastructures in Spain.

Author(s): Ji-Hye Baek, Sujin Kim, Seonghwan Choi, Jongyeob Park, Jihun Kim, Wonkeun Jo, Dongil Kim

Korea Astronomy and Space Science Institute; Korea Astronomy and Space Science Institute; Korea Astronomy and Space Science Institute; Korea Astronomy and Space Science Institute; Korea Astronomy and Space Science Institute; Chungnam National University; Ewha Womons University

Abstract: We have trained two models to detect solar events automatically using deep learning methods. To train the models, a dataset was required, however there were no existing deep learning datasets in the solar physics field. Too address this, we created DeepSDO event dataset using JPG data from 2010 to 2019 provided by the Korean Data Center for Solar Dynamics Observatory (KDC for SDO). We selected solar event such as coronal holes, sunspots, and prominences. This dataset is now publicly available on the KDC for SDO website (http://sdo.kasi.re.kr). We chose two representative deep learning object detection algorithms, Single Shot MulitiBox Detection (SSD) and Faster Region-based Convolutional Neural Network (R-CNN). Both models successfully detect solar events, and we are currently training them to detect additional solar events. Furthermore, we plan to create and provide a solar event catalog using our solar event auto detection models.

Author(s): Oleg Stepanyuk, Kamen Kozarev

Institute of Astronomy and National Astronomical Observatory of Bulgarian Academy of Sciences; Institute of Astronomy and National Astronomical Observatory of Bulgarian Academy of Sciences

Abstract: Over the last several decades a large amount of remote solar observational data has become available from ground-based and space-borne instruments. This has required the development of software approaches for automated characterisation of eruptive solar features. Most current solar feature detection and tracking algorithms have restricted applicability and complicated processing chains, while complexity in engineering machine learning (ML) training limits the use of data-driven approaches. Recently, we introduced Wavetrack – a general algorithmic wavelet decomposition-based method for smart characterization and tracking of solar eruptive features. In this work we present the natural evolution of this approach within the ML paradigm. We trained Convolutional Neural Network (CNN) image segmentation models using feature masks obtained by the Wavetrack code. We present the application of pre-trained models for segmentation of solar eruptive features and demonstrate their performance on a set of coronal eruptive events observed with the SDO/AIA instrument. We show how solar eruptive features obtained by our method can be used as ground truth for data-driven classification of radio instrument data.

Author(s): Francisco Iglesias, Florencia Cisterna, Diego LLloveras, Mariano Sanchez Toledo, Yasmin Machuca, Andrés Asensio Ramos, Franco Manini, Fernando Lopéz, Hebe Cremades

Max Planck Institute for Solar System Research and CONICET; GEHMe, University of Mendoza.; GEHMe, University of Mendoza and CONICET; GEHMe, University of Mendoza; GEHMe, University of Mendoza; Instituto de Astrofísica de Canarias; GEHMe, University of Mendoza and CONICET; GEHMe, University of Mendoza and CONICET; GEHMe, University of Mendoza and CONICET

Abstract: Coronal mass ejections (CMEs) are a major driver of space weather and thus can have important negative technological and social implications. Given our current inability to forecast the occurrence of a CME, it is crucial to assess their geoeffectiveness once they are ejected. Particularly relevant for this task is the reliable CME identification and correct assessment of its direction,  morphology and kinematics in coronagraph images. In the last decade, Deep Neural Networks (DNNs) have experienced enormous improvements in solving various machine-vision-related tasks, particularly excelling at image recognition and segmentation. One drawback when trying to use DNNs for CME segmentation in coronagraph images is that no large, curated dataset exists that can be used for supervised training. To mitigate this, we produced an ad-hoc synthetic dataset of CME coronagraph images by combining actual quiet (no CME) coronagraph background images with synthetic CMEs obtained using the Graduated Cylindrical Shell geometric model (GCS) and raytracing. We present results of a DNN-based model trained with our synthetic dataset to identify and segment the outer envelope of CMEs. This is done by fine tuning the pre-trained MaskR-CNN model, to produce a GCS-like mask of the CME present in a single differential coronagraph image. We compare the result of our method in terms of the main CME parameters (central position angle, angular width, velocity, etc.) with the results of three other wide-spread segmentation methods. This approach can also be used with any other model that allows generating synthetic CME brightness images and their corresponding mask, such as more realistic Magneto Hydrodynamic CME simulations.

Author(s): Francisco Iglesias, Mariano Sánchez Toledo, Andrés Asensio Ramos, Florencia Cisterna, Yasmin Machuca, Hebe Cremades

Max Planck Institute for Solar System Research and CONICET; GEHMe, University of Mendoza; Instituto de Astrofísica de Canarias, La Laguna, Tenerife, Spain; GEHMe, University of Mendoza; GEHMe, University of Mendoza; GEHMe, University of Mendoza and CONICET

Abstract: Coronal mass ejections (CMEs) are a major driver of space weather and can have important negative technological and social implications. They cannot be predicted and thus it is crucial to assess their geoeffectiveness once they are ejected. Particularly important is the reliable assessment of the CME 3D direction and kinematics. This is challenging because estimating the CME 3D morphology from the (up to) three 2D simultaneous coronagraphic views available is an ill-posed problem.  The parametric Graduated Cylindrical Shell geometric model (GCS) is heavily used by the community to help in the 3D reconstruction of CMEs. One important issue when applying the GCS model is that it is done manually and thus heavily relies on operator expertise. In recent years, Deep Neural Networks (DNN) have experienced enormous improvements in solving various machine-vision-related tasks, particularly excelling at image recognition and segmentation. One issue when trying to use these deep models for 3D reconstruction of CMEs is that no large, curated dataset exists that can be used for supervised training. To mitigate this, we produced a multi-viewpoint synthetic dataset of CME coronagraph images by combining actual quiet (no CME) coronagraph backgrounds with synthetic CMEs obtained with the GCS and raytracing. We present preliminary results of a DNN model that automatically reconstructs the 3D structure of the CME outer envelope from 3 simultaneous differential coronagraph images, acquired from different vantage points. This model is implemented by adding a fully connected linear head to a pre-trained deep convolutional backbone. The model is trained using our synthetic dataset to output the GCS model parameters that best fit the outer envelope of the CME observed in the input images.

Author(s): Seonghwan Choi, Sung-Hong Park, Ji-Hye Baek, Eunsu Park, Jeong-Heon Kim, Hwanhee Lee, Jongyeob Park, Harim Lee, Rok-Soon Kim, Chanwoo Kim, Yong-Jae Moon

Korea Astronomy and Space Science Institute; Korea Astronomy and Space Science Institute; Korea Astronomy and Space Science Institute; Korea Astronomy and Space Science Institute; Korea Astronomy and Space Science Institute; Korea Astronomy and Space Science Institute; Korea Astronomy and Space Science Institute; Kyung Hee University; Korea Astronomy and Space Science Institute; KAIST Academy; Kyung Hee University

Abstract: We are planning and promoting the SpaceAI program in Korea to support the application of artificial intelligence (AI) in the field of space weather. Raising research productivity may be the most valuable among all the advantages of AI. It can help discover more scientific knowledge and make science more efficient. The capabilities of AI itself are regression and classification, and they are being applied in various ways in the field of space science and technology. More specifically, AI is being applied to space science such as data processing and analysis, data-based discovery, large-scale surveys, and simulations, and to space technology such as autonomous operations, optimizing resource allocation, enhancing navigation and guidance, enhanced communication, astronaut support, and risk mitigation. Applying AI to research ultimately requires significant computing skills, although there are a variety of open tools that can be easily used. We can learn computing skills and conduct research at the individual research level, but we need to conduct research more quickly and achieve research results through collaboration with data scientists, programming experts, and AI experts. Additionally, hardware environments such as high performance computing systems, storage system for storing large amounts of data, and high-speed network must be provided. Many science and technology researchers are trying to apply AI to their research, but we have found that it takes too much time to get results or they give up altogether. The SpaceAI program has been conceived with the primary objective of reducing the entry barrier for researchers seeking to leverage artificial intelligence (AI) in the domain of space science and technology. This initiative aims to foster collaboration among experts across diverse fields and provide access to the abundant computing resources of various institutions. Additionally, the program endeavors to endorse the dissemination of intellectual properties, including patents and software, as well as papers. Through these concerted efforts, we are expecting that SpaceAI not only cultivate a vibrant research culture but also yield significant and valuable contributions to the field.

Author(s): Hassan Nooreldeen, Ayman Ahmed, Karim Garad, Abdallah Shaker, Amira Hussien

Egyptian Space Agency (EgSA); Egyptian Space Agency (EgSA); Egyptian Space Agency (EgSA); Egyptian Space Agency (EgSA); Egyptian Space Agency (EgSA)

Abstract: North Africa faces challenges in accessing ionospheric ground observational data, hindering accurate forecasts. This study aims to bridge this gap by developing an accurate ionospheric prediction model to anticipate the response to space weather activities. Ground-based ionospheric data derived from Global Navigation Satellite System (GNSS) observations were utilized as the basis for our model. Leveraging data from 12 GNSS ground stations, we focused on midlatitude regions over North Africa, covering geomagnetic latitudes at 35°N, 30°N, and 25°N. The dataset spanned five years and included essential geomagnetic indices—Dst, F10.7, and KP—alongside solar wind parameters to enhance prediction accuracy. Vertical total electron content (VTEC) was employed as input for the DNN to accurately represent the ionospheric state. A significant portion of the research was dedicated to data preparation and pre-analysis, which proved to be the most time-consuming step. Various Deep Neural Network (DNN) architectures were tested to identify the model with the lowest Root Mean Square Error (RMSE). The selected DNN model demonstrated a substantial improvement in forecasting ionospheric VTEC changes driven by space weather activities. By integrating ground-based GNSS data with geomagnetic indices and solar wind parameters, our DNN model effectively trained on complex time-series datasets, yielding precise forecasts.

Author(s): Ticiano Torres Peralta, Maria Graciela Molina, Hernan Asorey, Ivan Sidelnik, Antonio J. Rubio-Montero, Rafael Mayo-Garcia, Alvaro Taboada, Sergio Dasso, Luis Otiniano, for the LAGO Collaboration

FACET-UNT/CONICET; FACET-UNT/CONICET/INGV; Departamento de fısica de neutrones, Centro Atomico Bariloche, Comision Nacional de Energıa Atomica (CNEA), Argentina; Instituto de Tecnologıas en Deteccion y Astropartıculas (ITeDA), Argentina; Centro de Investigaciones Energeticas Medioambientales y Tecnologicas (CIEMAT), Spain; Centro de Investigaciones Energeticas Medioambientales y Tecnologicas (CIEMAT), Spain; Instituto de Tecnologıas en Deteccion y Astropartıculas (ITeDA), Argentina; Instituto de Astronomıa y Fısica del Espacio (IAFE), Argentina; Comision Nacional de Investigacion y Desarrollo Aeroespacial (CONIDA), Peru; The LAGO Collaboration, see the complete list of authors and institutions at https://lagoproject.net/collab.htmlhttps://lagoproject.net/collab.html

Abstract: The Latin American Giant Observatory (LAGO) is a ground-based, extended observatory network of Water Cherenkov Detectors (WCD) for measuring the secondary particle flux from astroparticles interacting with Earth’s atmosphere. These secondary particles are typically grouped into three distinct components: the electromagnetic, muonic, and hadronic components. The WCD response to a given particle can be observed in the so-called signal- charged histogram. The time variability of this signal flux at the WCD can be correlated with space weather-related phenomena on time scales from hours to weeks.
In this work, we implemented OPTICS, a density-based clustering algorithm, to identify patterns in data collected by a single WCD. We implemented the algorithm in two ways: a) using actual data from a WCD at the LAGO site in Bariloche, Argentina, and b) validating our initial findings using a synthetic (simulated) dataset. This simulated dataset is the result of combining the outputs of the ARTI and MEIGA simulation frameworks. This dataset simulates the expected WCD signals produced by the flux of secondary particles during one day at the LAGO site in Bariloche.
The results show that we can categorize eight candidate groups in a robust manner using actual data. We corroborated these findings by obtaining similar results with the synthetic data. Our analysis demonstrates that applying our methodology can accurately identify the originating particle with a high degree of confidence on a single-pulse basis, highlighting its precision and reliability

Author(s): Sanjay Gosain

National Solar Observatory

Abstract: We apply machine learning methods to detect chromospheric features on solar H-alpha fulldisk images in near real-time from the GONG network. In particular we discuss the algorithms to detect the filaments on the disk and then using deep learning methods detect activated filaments that show pre-eruption dynamics such as slow rise and activated oscillations. The location of the filaments on the disk and pre-eruption dynamics can be used to predict the probability of Earth directed coronal mass ejections. Detection of such events along with their coronal magnetic field models can be used to make better estimates of geoeffectiveness of the associated CMEs.

Author(s): Ingo Michaelis, Jan Rauberg, Martin Rother, Guram Kervalishvili, Monika Korte

GFZ German Research Centre For Geosciences; GFZ German Research Centre For Geosciences; GFZ German Research Centre For Geosciences; GFZ German Research Centre For Geosciences; GFZ German Research Centre For Geosciences

Abstract: Magnetic field data from satellite missions play a significant role in characterizing and understanding space weather conditions. Satellite magnetic field observations from different altitudes and local times are necessary to disentangle the complex processes contributing to each observation of Earth’s magnetic field and study the various individual processes. While the dedicated magnetic field satellite missions give good global data coverage at first sight, the coverage is still sparse if simultaneous observations from several different altitudes and with a good local time coverage are desired. Moreover, gaps between dedicated magnetic field satellite missions, such as between the CHAMP and Swarm missions from 2010 to 2013, exist and might occur again in the future. Many satellites non-dedicated for magnetic field measurements carry so-called platform magnetometers (PlatMags) that are part of the attitude and orbit control system (AOCS). These satellites have a variety of mission goals and the PlatMags are additional instrumentation for navigational use.
Using Machine Learning tools and additional satellite telemetry data we can remove artificial disturbances from the satellite magnetometers and calibrate PlagMags for scientific use.

Author(s): Eunsu Park, Harim Lee, Seonghwan Choi, Yong-Jae Moon

Korea Astronomy and Space Science Institute; Kyung Hee University; Korea Astronomy and Space Science Institute; Kyung Hee University

Abstract: In this study, we propose a simple deep learning model for the removal of noise from solar magnetograms. In our previous study, we successfully applied an image-to-image translation model to translate a single frame solar magnetogram to a 21-frame-stacked magnetogram, effectively reducing noise. However, the previous method had two limitations: the requirement to prepare a substantial number of pairs of original magnetograms and target denoised magnetograms for training the deep learning model, and the performance dependency on the quality of the produced denoised magnetogram. In this work, we aim to introduce a novel and simplified deep learning model and methodology to address these issues. Our proposed model can be trained without target magnetograms (denoised magnetograms) and demonstrates denoising performance comparable to or better than the previous model. Furthermore, this methodology can be applied to various magnetograms observed from different observatories.

Author(s): Harim Lee, Eunsu Park, Jihyeon Son, Yong-Jae Moon

Kyung Hee University; Korea Astronomy and Space Science Institute; Kyung Hee University; Kyung Hee University

Abstract: Solar active regions are important for solar physics research and have been used as an input for various space weather forecasting models. In this study, we are developing a prediction model for solar active region evolution using video prediction methods based on deep learning. For the dataset, we use magnetograms observed by Solar Dynamics Observatory (SDO) / Helioseismic and Magnetic Imager (HMI). For magnetograms we use the Cylindrical Equal Area (CEA) projected HMI Active Region Patches (HARPs) data provided by JSOC, which have an advantage that they have the same area per pixel, without any additional data pre-processing. We consider 720 HARPs within ±30 degrees of the solar central meridian and with HARPs of 512×512 size from May 2010 to May 2023 (about 1 million images at 12 minute intervals). For the deep learning models, we consider several spatiotemporal prediction models, such as ConvLSTM and Top-tier video prediction models. Our models are designed to predict the change of active regions over the next 12 hours. Our study is a new attempt to predict the evolution of solar active regions using deep learning and is meaningful in providing useful information on solar activity forecasts.

Author(s): Silvia Kostárová, Adam Majirský, Samuel Amrich, Šimon Mackovjak

IEP Slovak Academy of Sciences; FEI Technical University of Košice; IEP Slovak Academy of Sciences; IEP Slovak Academy of Sciences

Abstract: The Vigil mission, planned by the European Space Agency, will monitor space weather activity, providing warnings about threats coming from the Sun earlier than current missions allow. Until the data from Vigil is available, we can use measurements from other missions of the heliophysics fleet and consider their similarities to Vigil. Specifically, we focus on the most extreme events of the last 30 years. Our goal is to answer the following questions: Would it be possible to predict the occurrence of extreme space weather events if the Vigil mission was in operation at that time? Is it beneficial to use Machine Learning (ML) techniques for these predictions? In our contribution, the concept of Vigil-like instruments is introduced to describe existing space-based instruments with capabilities similar to those planned for Vigil. Moreover, methods for the community to access and process this data is presented. Finally, the poster is focusing on our major ML tasks, including how the segmentation of coronal structures can be useful and what is the potential of Vigil-like in-situ measurements in the early prediction of extreme events. These tasks are under development within the ongoing study supported by ESA through the RPA program in Slovakia.

Author(s): Bibhuti Kumar Jha, Chetraj Pandey, Opal Issan, India Jackson, Mike Heyns, Raman Mukundan, Vishal Upendran, Banafsheh Ferdousi, Panos Tigas, Hira Saleem, Alex Lavin

SWRI; Georgia State University; University of California San Diego; Georgia State University; Imperial College London; University of New Hampshire; Lockheed Martin Solar and Astrophysics Laboratory (LMSAL); AFRL; Isomorphic; University of New South Wales; Pasteur Labs & ISI

Abstract: With increasing reliance on modern infrastructure, forecasting space weather has become critical to societal resilience. For these forecasts to be actionable to end-users, they need near real-time inference, provide localised context, allow enough lead time for mitigation strategies to be implemented, and finally include an associated estimate of prediction uncertainty. Within the modern paradigm of data driven models and ML/AI approaches, there is the ability to deliver on these forecasting needs and complement existing physics based simulations. Building on previous work [Upendran et al. 2020, FDL-X 2023], we present a generalised pipeline for ground geomagnetic field perturbation forecasts globally. Two separate regimes are included, the first using high fidelity real-time in-situ solar wind measurements for modelling, and the second using a coupled low fidelity ambient solar wind model that provides increased lead time. Further additions include uncertainty estimation within both regimes and the enhancement of the individual models. The pipeline presented makes use of continual learning and online inference, providing a roadmap for other ML/AI to transition to operations and remain relevant with future data streams and new events.
This work is the research product of FDL-X Heliolab, a public/private partnership between NASA, Trillium Technologies Inc and commercial AI partners Google Cloud, NVIDIA and Pasteur Labs & ISI, developing open science for all Humankind.

Author(s): Daniel Holmberg, Ivan Zaitsev, Markku Alho, Fanni Franssila, Ioanna Bouri, Minna Palmroth, Teemu Roos

University of Helsinki; University of Helsinki; University of Helsinki; University of Helsinki; University of Helsinki; University of Helsinki; University of Helsinki

Abstract: Space weather forecasting is crucial for mitigating the risks posed by solar wind interactions with Earth’s magnetosphere, which can disrupt power grids, satellite operations, and communication networks. Traditional forecasting methods rely on magnetohydrodynamic (MHD) simulations. However, these models are limited by their simplified fluid dynamics approach. Hybrid-Vlasov simulations offer a more detailed representation of magnetospheric dynamics by capturing both fluid and ion kinetic processes, providing a significant improvement over MHD simulations to model non-linear multiscale events such as e.g. magnetic reconnection. Despite their increased accuracy, these simulations are computationally prohibitive for real-time applications, often requiring orders of magnitude more computational resources.
To address these challenges, we propose a novel machine learning approach to emulate hybrid-Vlasov simulations, that requires a fraction of the compute to run compared to numerical simulation. Our study demonstrates the feasibility of this approach by training an autoregressive hierarchical graph neural network on 2D simulation data generated by Vlasiator. This machine learning-based emulator not only achieves substantial speedups over traditional numerical simulations, but also successfully identifies potential sites for future magnetic reconnection events.
Looking ahead, we plan to explore the impact of various boundary conditions on the learned forecast and enhance the model’s physical realism by incorporating constraints, such as encouraging zero divergence in the predicted magnetic fields. To facilitate further research and innovation, we intend to release an open data benchmark for machine learning practitioners. This benchmark will assess model generalization across different initial conditions, promoting the development of more efficient and accurate forecasting models.

Author(s): Fanni Franssila, Ioanna Bouri, Markku Alho, Daniel Holmberg, Minna Palmroth, Teemu Roos

University of Helsinki; University of Helsinki; University of Helsinki; University of Helsinki; University of Helsinki; University of Helsinki

Abstract: Magnetic reconnection is a fundamental plasma physics phenomenon where energy stored in a magnetic field is released in the form of high-velocity plasma jets. When occurring in near-Earth space, the resulting jets carry the risk of damaging the technological infrastructure on Earth or on orbit. Similarly, reconnections can happen in fusion reactors, where they can disrupt the fusion process by breaking plasma confinement. To prevent such risks, it is imperative to develop efficient and accurate methods for detecting potential sites for oncoming magnetic reconnections.
While magnetic reconnection is relatively well-understood in two dimensions, the phenomenon remains challenging to characterize in 3-dimensional (3D) settings. In 3D magnetic fields, reconnection sites are typically located around separators, which are special field lines derived from the global topology of the magnetic field. Therefore, in order to identify or predict 3D reconnections, it is natural to compute the separators, and to consider them as potential reconnection sites. However, the 3D magnetic fields considered in real-world settings, such as the Earth’s magnetotail, typically exhibit highly complex topological structures. Processing such 3D topologies can lead to significantly heavy computational loads. This, along with the current lack of understanding of 3D reconnection, makes it essential to develop new methods that are not only accurate but also computationally efficient in detecting separators.
Current approaches to locating the separators of a 3D magnetic field include, for example, methods that trace separators from other topological elements, such as the Magnetic Skeleton Analysis Tools (MSAT) package, and methods that approximate separators based on the local topology of magnetic field lines. However, these methods tend to impose a significant trade-off between accuracy and computation time on large-scale data. To this end, we present  a machine learning (ML) approach to detecting separators with the aim of maintaining accuracy while offering an increase in computational efficiency.
In our work, we develop an ML surrogate model for the separator computation method implemented in the MSAT package, which is one of the more accurate but relatively slow existing approaches. We use our ML model to speed up the detection of separators. The model is trained to take magnetic field data as input, and produce the separators as output using MSAT separators as the ground truth when training the model.
We train and test our model on simulated data of the Earth’s magnetotail, produced using Vlasiator – a 3D hybrid-Vlasov plasma simulation of near-Earth space. This allows us to validate our model on complex, large-scale data. As a result, our model approximates the MSAT separator output on complex data with significant computational speed-up.
Keywords: magnetic reconnection, machine learning, magnetic fields, topology.

Author(s): Ioanna Bouri, Fanni Franssila, Markku Alho, Giulia Cozzani, Ivan Zaitsev, Daniel Holmberg, Minna Palmroth, Teemu Roos

University of Helsinki; University of Helsinki; University of Helsinki; University of Helsinki; University of Helsinki; University of Helsinki; University of Helsinki; University of Helsinki

Abstract: Magnetic reconnection in plasmas transforms magnetic energy into kinetic and thermal energy while the magnetic field undergoes a topological reconfiguration. The phenomenon can be encountered in various settings, from the Earth’s magnetosphere and solar corona to fusion plasmas in tokamaks. In space weather, magnetic reconnection can have severe real-world implications – it is linked to eruptive events like solar flares and coronal mass ejections, which can cause power outages, damage technological infrastructure, and disrupt satellite communications.
While the dynamics of magnetic reconnection are well-understood in two dimensions, the added complexity of a three-dimensional (3D) setting introduces substantial challenges for effectively characterizing and understanding the phenomenon. Advancing our understanding of 3D magnetic reconnection can improve modeling techniques and simulations, leading to better mitigation of active space weather events. Interpretable machine learning offers new possibilities for deepening our understanding of 3D magnetic reconnection, allowing us to validate existing models and uncover novel features within this complex physical process.
The 3D topology of magnetic fields can be described using magnetic skeletons, which include topological elements such as magnetic nulls, separatrices, and separators, which constitute potential reconnection sites. However, the detection of separators can be computationally expensive, and thus, approximations, such as quasi-2D methods, are often used. Such approaches are based on our currently limited physical understanding of the phenomenon, whereas purely data-driven methods for feature discovery are not as constrained by such limitations.
When a convolutional neural network (CNN) is trained to perform a specific classification task, gradient-weighted class activation mapping (GradCAM) is a technique that can be used to identify the regions in the original input data that contributed the most to the predictive performance by measuring the gradient of the classification score with respect to the CNN’s convolutional features. We have developed a framework of different 3D CNN variations, and for each of them, we generated and examined the corresponding GradCAM heatmaps. The CNNs are trained using 3D plasma simulations of the near-Earth magnetosphere created by Vlasiator, a simulator based on Vlasov theory. The networks are trained to classify active reconnection sites in 3D on a cell-by-cell basis, focusing on areas where particle acceleration and plasma jet reversals occur near the separators.
Initial findings indicate that the GradCAM heatmaps:
(a) Confirm the importance of separators in detecting reconnection sites. This fact verifies our 3D GradCAM method since the model targets were constructed using the quasi-2D approximations of the magnetic X-lines.
(b) Show that GradCAM heatmaps can detect active reconnection sites downstream of the Vx zero-crossing contour.
(c) Validate that GradCAM heatmaps can capture other phenomena related to magnetic reconnection, such as the bursty bulk flow (BBF) regions.
Finally, it was unclear why some prominent regions in the GradCAM heatmaps were presented as such. No obvious combination of the model’s input features seemed to suggest they were active reconnection sites. However, we discovered that (d) these regions became increasingly active during the following timesteps of the simulation (cadence=1s), suggesting that GradCAM could be picking up features useful for 3D reconnection forecasting.
Keywords: Deep learning, 3D magnetic reconnection, Interpretability

Author(s): Gautier Nguyen, Antoine Brunet, Maria Tahtouh, Guillerme Bernoux, Sebastien Bourdarie, Vincent Maget, Ingmar Sandberg

DPHY ; ONERA, Université de Toulouse ; Toulouse, France; DPHY ; ONERA, Université de Toulouse ; Toulouse, France; DPHY ; ONERA, Université de Toulouse ; Toulouse, France; DPHY ; ONERA, Université de Toulouse ; Toulouse, France; DPHY ; ONERA, Université de Toulouse ; Toulouse, France; DPHY ; ONERA, Université de Toulouse ; Toulouse, France; Space Applications & Research Consultancy (SPARC), Aiolou St. 73, 10551 Athens, Greece

Abstract: Reduced Order Models (ROMs) are fast, powerful alternatives to the more computationally expensive physics-based models. This for their capacity to extract underlying non-linear features of high dimensional data into a reduced set of latent variables,
From a space weather perspective, recent ROM development typically include thermosphere (Licata&Mehta 2018, Nateghi&Manzi 2022) and ring current modeling (Cruz et al. 2024).
The ever-growing amount of in-situ data is a key asset for the development of Deep Learning based dimensionality reduction techniques such as non-linear Auto-Encoders. Nevertheless, their application to physical systems are often limited by their non interpretable or entangled latent variables and the necessary blindfold determination of their optimal latent dimension.
This work addresses these issues in the specific context of the Earth electron radiation belt modeling through the use of a β-Variational Auto-Encoder, a specific type of auto-encoder that ensures both disentangled, thus more interpretable, and information intensive latent variables.
Applied on long-term simulations of the electron belts between 2014 and 2019, we show that the high dimensional complexity of the radiation belt dynamics can actually be reduced to a small number of latent variables, each of them exhibiting strong correlations with a small subset of solar and geomagnetic external parameters. This constitutes the first step towards the design of a reduced order model of the Earth electron radiation belts parametrized by geomagnetic and solar proxies only.
This work was supported by both the “Event-Based Electron Belt Radiation Storm Environments Modelling” Activity led by the Space Applications & Research Consultancy (SPARC) under ESA Contract 4000141351/23/UK/EG and by the “Global Radiation Belt Prototype for LEO Constellations” Activity led by ONERA under ESA Contract 4000137689/22/NL/CRS.

Author(s): Philippe Vong, Laurent Dolla, Alexandros Koukras, Jacques Gustin, Jorge Amaya, Ekaterina Dineva, Giovanni Lapenta

Royal Observatory of Belgium; Royal Observatory of Belgium; Columbia Astrophysics Laboratory, Columbia University; Royal Observatory of Belgium; Center for Mathematical Plasma Astrophysics, KULeuven; Center for Mathematical Plasma Astrophysics, KULeuven; Center for Mathematical Plasma Astrophysics, KULeuven

Abstract: The spatial extension of active regions of the Sun (hence the size of their associated images) can strongly vary from one active region to another. This size inhomogeneity is a problem when studying solar flares with Convolutional Neural Networks (CNNs), as CNNs generally use input images of a fixed size. Several processes can be performed to produce a data set with homogeneous-sized data, such as coarse resizing or cropping/padding of raw images. Unfortunately, key features can be lost or distorted beyond recognition during these processes. This can lead to a deterioration of the ability of CNNs to classify flares of different soft X-ray classes, especially those from active regions with structures of great complexity.
This study aims to implement and test a CNN architecture that retains features of characteristic scales as fine as the original resolution of the input images. This study compares the performance of two convolutional neural network models for solar flare prediction: the first one is a traditional CNN with convolution layers, batch normalization layers, max-pooling layers, and resized input whereas the other implements a spatial pyramid pooling (SPP) layer instead of a max pooling layer before the flatten-layer and without any input resizing. Both are trained on the SHARP Line-of-sight magnetogram database from 2010-05 to 2021-08 and use only images within 45° of the central meridian of the Sun. We also study two cases of binary classification: in the first case, our model has to distinguish active regions producing flares in less than 24h of class ≥C1.0 from active regions producing flares in more than 24h or never; in the second case, it has to distinguish active regions producing flares in less than 24h of class ≥M1.0 from active regions producing flares in more than 24h or never, or flares in less than 24h but of class <M1.0.
Our model implementing an SPP layer predicts flares ≥C1.0 within 24 hours more accurately than the traditional CNN model with a 10% increase in recall. We proved the higher efficiency of a CNN model that includes a spatial pyramid pooling layer in predicting solar flares.
However, its performances degrade sharply when trained to classify images of ≥M1.0 flares. The degradation of prediction performance of this model when the images of active regions producing a C-class flare are classified as negative may be attributed to its success in identifying features that appear in active regions only a few hours before the flare, independently of their soft X-ray class. The development of explainable artificial intelligence tools adapted to this architecture in future projects will be interesting for the study of solar flare-triggering mechanisms.

Author(s): Armando Collado Villaverde, Pablo Muñoz Martínez, Consuelo Cid Tortuero

Universidad de Alcalá; Universidad de Alcalá; Universidad de Alcalá

Abstract: Geomagnetic storms pose substantial risks to modern technological infrastructure, making accurate forecasting of these events critical for mitigating their impacts on essential systems. In the current landscape of space weather research, significant emphasis has been placed on the forecasting of global geomagnetic indices such as Kp, Dst, SYM-H, and Hp30. While these global indices provide valuable insights, the scientific community has paid comparatively less attention to local geomagnetic indices, despite their potential importance. This gap in research is particularly evident considering the geomagnetic storm of May 2024, when auroras were observed at unusually low latitudes but not uniformly across all longitudes. This event highlights the necessity of focusing on local geomagnetic disturbances, as they can provide more precise information about regional space weather effects.
Considering that, our work emphasizes forecasting local geomagnetic indices, specifically the Local Disturbance Index (LDi), using data from four observatories distributed across different longitudes. This regional approach allows for a more precise forecast of geomagnetic disturbances and their potential effects at specific geographic locations. To this end, we have developed a neural network that forecasts the LDi by using historical geomagnetic data along with real-time solar wind measurements from ACE. The input data for the model includes previous values of the LDi for the specific observatory, solar wind parameters such as the interplanetary magnetic field, and proton density, speed and temperature. To take into account the geographical nature of the forecast, we incorporated additional parameters, namely the magnetic local time (MLT) and the latitude of each target observatory. These factors are critical in understanding the diurnal and latitudinal variations in geomagnetic activity, which are neglected in global indices. Furthermore, we have included a prediction interval with 90% confidence, providing a measure of the uncertainty in our forecasts, which is particularly important in operational space weather forecasting.
The development of this neural network model addresses the limited focus on local indices in the existing literature and offers a more targeted approach of space weather to the users’ needs. By integrating real-time solar wind data and observatory-specific characteristics, our work offers a prediction of local geomagnetic disturbances. This advancement is crucial for better understanding the regional impact of geomagnetic storms, allowing to assess local geomagnetic disturbances that can be masked in global indices.
Through the application of neural networks and the inclusion of solar wind data alongside local observatories, we propose a tool for predicting geomagnetic disturbances at a regional level. This will enhance the resilience of critical systems for local areas where magnetometers are available to compute and forecast the LDi, providing warnings tailored to the region.