Session CD1 - Artificial intelligence in the service of space weather

Elena Popova (Centro de Investigación de Astronomía, Universidad Bernardo O’Higgins, Chile), Robertus Erdelyi (University of Sheffield, Sheffield, UK), Marianna Korsos (Aberystwyth University, Aberystwyth, UK), Giovanni Lapenta (KU Leuven, Belgium)

Artificial intelligence is taking paramount importance in a wide range of applications from engineering to space physics. A particularly interesting area is big data and its associated applications. In the last few years, machine learning techniques have proven capable of forecasting space weather events with a much higher accuracy with respect to long-used traditional empirical and physics-based models.The direction of space weather uses big data, as an example, that are hard to handle with traditional methodology. It is hard to imagine the future of space weather without machine learning because more consideration is being given to the issues of reliability, uncertainty, and trustworthiness of machine learning models. O)n the practical sifde, the forecasts of the various physical processes are especially timely given the recent technology developments and the expansion of our technosphere. This session encourages submissions addressing the latest advances in the application of artificial intelligence and their application to space weather. Contributions are welcome from all areas of space weather that focus on the application of artificial intelligence, including forecasting various processes and analyzing satellite or ground-based data.

Poster Viewing
Monday October 24, 09:00 - 14:00, Poster Area

Talks
Monday October 24, 13:45 - 15:00, Earth Hall
Monday October 24, 16:00 - 17:00, Earth Hall
Tuesday October 25, 17:00 - 18:00, Earth Hall

Click here to toggle abstract display in the schedule

Talks : Time schedule

Monday October 24, 13:45 - 15:00, Earth Hall

13:45	Automatic Detection of Interplanetary Coronal Mass Ejections in Solar Wind In Situ Data	Rüdisser, H et al.	Oral
	Hannah T. Rüdisser[1], Andreas Windisch[2,3,4,5] , Ute V. Amerstorfer[1], Christian Möstl[1], Rachel L. Bailey[6], Tanja Amerstorfer[1,7]
	[1]Austrian Space Weather Office, Zentralanstalt für Meteorologie und Geodynamik, Graz, Austria; [2]Know Center, Graz, Austria; [3]Institute of Interactive Systems and Data Science, Graz University of Technology, Graz, Austria; [4]Department of Physics, Washington University in St. Louis, MO 63130, USA; [5]RL Community, AI AUSTRIA, Vienna, Austria; [6]Conrad Observatory, Zentralanstalt für Meteorologie und Geodynamik, Graz, Austria; [7]Space Research Institute, Austrian Academy of Sciences, Graz, Austria
	Interplanetary coronal mass ejections (ICMEs) are one of the main drivers for space weather disturbances. In the past, different approaches have been used to automatically detect events in existing time series resulting from solar wind in situ observations. However, accurate and fast detection still remains a challenge when facing the large amount of data from different instruments. For the automatic detection of ICMEs we propose a pipeline using a method that has recently proven successful in medical image segmentation. Comparing it to an existing method, we find that while achieving similar results, our model outperforms the baseline regarding training time by a factor of approximately 20, thus making it more applicable for other datasets. The method has been tested on in situ data from the Wind spacecraft between 1997 and 2015 with a True Skill Statistic (TSS) of 0.64. Out of the 640 ICMEs, 466 were detected correctly by our algorithm, producing a total of 254 False Positives. Additionally, it produced reasonable results on datasets with fewer features and smaller training sets from Wind, STEREO-A and STEREO-B with True Skill Statistics of 0.56, 0.57 and 0.53, respectively. Our pipeline manages to find the start of an ICME with a mean absolute error (MAE) of around 2 hours and 56 minutes, and the end time with a MAE of 3 hours and 20 minutes. The relatively fast training allows straightforward tuning of hyperparameters and could therefore easily be used to detect other structures and phenomena in solar wind data, such as corotating interaction regions.
14:00	Advanced Image Preprocessing and Feature Tracking for Remote CME Characterization with Convolutional Neural Network	Stepanyuk, O et al.	Oral
	Oleg Stepanyuk, Kamen Kozarev
	Institute of Astronomy and National Astronomical Observatory of Bulgarian Academy of Sciences
	Coronal Mass Ejections (CMEs) influence the interplanetary environment over vast distances in the solar system by injecting huge clouds of fast solar plasma and energetic particles (SEPs). A number of fundamental questions remain about how SEPs are produced, but current understanding points to CME-driven shocks and compressions in the solar corona. At the same time, unprecedented remote (AIA, LOFAR, MWA) and in situ (Parker Solar Probe, Solar Orbiter) solar observations are becoming available to constrain existing theories, nevertheless, reliable training sets for classification, segmentation and tracking of CME-related phenomena with CNN models are still missing. Recently (Stepanyuk et.al, J. Space Weather Space Clim. Vol 12, 20 (2022)), we have demonstrated the method and the software (https://gitlab.com/iahelio/mosaiics/wavetrack) for smart characterization and tracking of solar eruptive features based on the a-trous wavelet decomposition technique, intensity rankings and a set of filtering techniques. The typical use of СNN's is on classification tasks, where the output to an image is a single class label. However, the desired output should include localization. More-over, obtaining tens of thousands of verified training images was out of reach, as fast and reliable method of obtaining features masks was not present. The task is still somewhat problematic even with automated algorithmic segmentation tool such as Wavetrack, taking into account the desired amount of images within typical CNN-traning routine. We adopt U-Net for Solar Eruptive Feature Extraction and Characterization. U-Net is a convolutional neural network that was developed for biomedical image segmentation at the Computer Science Department of the University of Freiburg (Ronneberger O, et. al (2015). This network and training strategy relies on the strong use of data augmentation to use the available annotated samples more efficiently and can be trained end-to-end from a very limited set of images, while feature enginering allows to improve this approach even further by expanding available training sets. Here we present pre-trained models and demonstrate data-driven characterization and tracking of solar eruptive features on a set of CME-event related datasets obtained from SDO/AIA telescope.
14:15	Probing the coronal magnetic field with physics informed neural networks	Jarolim, R et al.	Oral
	Robert Jarolim[1], Julia Thalmann[1], Astrid Veronig[1,2], Tatiana Podladchikova[3]
	[1]University of Graz, [2]Kanzelhöhe Observatory for Solar and Environmental Research, [3] Skolkovo Institute of Science and Technology
	While the photospheric magnetic field of our Sun is routinely measured, its extent into the upper solar atmosphere (the corona) remains elusive. The 3D distribution of the coronal magnetic field is essential to understand the genesis and initiation of solar eruptions and to predict the occurrence of high-energy events from our Sun. We present a novel approach for coronal magnetic field extrapolation using physics informed neural networks. The neural network is optimized to match observations of the photospheric magnetic field vector at the bottom-boundary, while simultaneously satisfying the force-free and divergence-free equations in the entire simulation volume. We demonstrate that our method can account for noisy data and deviates from the physical model where the force-free magnetic field assumption cannot be satisfied. We utilize meta-learning concepts to simulate the evolution of the active region 11158. Our simulation of 5 days of observations at full cadence, requires less than 13 hours of total computation time. The derived evolution of the free magnetic energy and helicity in the active region, shows that our model captures flare signatures, and that the depletion of free magnetic energy spatially aligns with the observed EUV emission. Our method provides the ability to perform magnetic field extrapolations in quasi real-time, which can be used for space weather monitoring, studying pre-eruptive structures and as initial condition for MHD simulations.
14:30	Using Neural Networks to improve the performance and forecasting skill of a solar wind model	Barros, F et al.	Oral
	Filipa S. Barros[1], Rui F. Pinto[2], J. J. G. Lima[3], André Restivo[1]
	[1] LIACC & FEUP, Portugal, [2] IRAP/CNRS & LDE3/CEA Saclay, France, [3] DFA, FCUP & IA/CAUP, University of Porto,Portugal
	We investigate whether and how machine learning methods can be applied to the solar wind model MULTI-VP noth to improve the quality of its solutions and its computational performance. MULTI-VP computes vast collections of individual solar wind streams, predicting the wind speed, density, temperature and magnetic field amplitude for each streams all the way from the surface of the Sun to about 15% of the Sun-Earth distance. Its solutions are then fed into heliospheric propagation models that determine the state of the background solar wind at Earth or at other solar system bodies with a few-day lead time. The performance of this kind of numerical simulation depends strongly on the quality of the initial guesses provided. We have applied a sequential model RNN with eight hidden dense layers and a dropout of 0.2 and 52 nodes was trained with 6000 wind profiles for 500 epochs, and nuilt a system that proposes initial solar wind profiles tailored to each stream considered. Our method resulted, for now, on a simulation tie speedup of about 13% on average (with a paired t-test p-value of 0.013) and also (perhaps more importantly) to more stable numerical behaviour (with lower amplitude transients). We plan on implementing this new tool on the automated SWiFT/MULTI-VP solar wind forecasting pipeline in the near future.

Monday October 24, 16:00 - 17:00, Earth Hall

16:00	Probabilistic ensemble learning for flare forecasting and value-weighted assessment	Guastavino, S et al.	Oral
	Sabrina Guastavino[1], Francesco Marchetti[2], Michele Piana[1], Federico Benvenuto[1], Cristina Campi[1]
	[1] MIDA group, Dipartimento di Matematica, Università di Genova, Italy, [2] Dipartimento di Matematica, Università di Padova, Italy
	There are three technical aspects that may have significant impacts on the forecasting effectiveness of machine learning for flare prediction. First, the optimization of the network parameters depends on an appropriate choice for the loss function in the training phase; second the validation step requires to implement a stopping rule that allows the selection of the best epoch; finally, the assessment of the prediction power of the algorithm should account for the dynamical nature of the flare forecasting problem. This methodological talk shows that probabilistic score-oriented loss (SOL) functions allow an optimization of the network that automatically accounts for the kind of skill score used for the prediction assessment; then, it describes how ensemble learning can be used in the validation phase to realize epoch selection; and, finally, the talk introduces value-weighted skill scores for flare forecasting that give greater importance to the values of the prediction than to its quality. We will finally point out that these aspects are crucial both in feature-based machine learning and image-based deep learning.
16:15	A prototype for a PCA-NN model for TEC with space weather parameters as predictors: selection of a NN algorithm and a set of predictors	Morozova, A et al.	Oral
	Anna Morozova, Ricardo Gafeira, Teresa Barata, Tatiana Barlyaeva
	Instituto de Astrofísica e Ciências do Espaço, University of Coimbra, OGAUC, Coimbra, Portugal
	A PCA-NN model for the total electron content (TEC) for the midlatitudinal region (Iberian Peninsula) is based on the decomposition of the observed TEC series using the principal component analysis (PCA) and the reconstruction of the TEC parameters (the daily mean TEC and amplitudes of the daily PCA modes) with neural networks (NN). Different NN algorithms (feedforwards with weight backpropagation, convolutional, recurrent) were tested and the best sets of predictors (space weather parameters) were defined for each of the TEC parameters and for each of the NN types. Based on these results a prototype for a PCA-NN model for TEC is proposed.
16:30	Temporal Convolutional Network for Local Forecast of Precipitated Electron Energy Flux	Bouriat, S et al.	Oral
	Simon Bouriat[1], Mathieu Barthélémy[2], Jocelyn Chanussot[3]
	[1]IPAG, Gipsa-Lab, SpaceAble, [2]IPAG, CSUG, [3]Gipsa-Lab
	The \textit{new space} industry brought fresh challenges to the space weather community as new satellites often appear more vulnerable to the natural environment. A recent incident involving Starlink satellites’ loss caused by a geomagnetic storm shows the need for a better understanding of space weather-related hazards, especially in LEO where new constellations are about to be brought. One danger of solar activity is \textit{spacecraft charging}: the accumulation of charged particles on and inside satellites, triggering electrostatic discharge. Only a few satellites measure them, and there still is a lack of models to characterize them. The exponential growth in the use of AI techniques has shown tremendous results and is very promising to answer this need. In this context, our study aims at presenting the first results of using a temporal convolutional network (TCN) to nowcast and forecast precipitated electron energy flux as measured by the satellites of the Defense Meteorological Satellite Program. Our inputs are solar wind data propagated at the bow shock nose (BSN) from OMNIWeb platform, and our outputs are time-series of electron energy flux for a given Magnetic Local Time and Magnetic Latitude. Using high-resolution OMNIWeb data to forecast close-to-Earth variables is not new and has been done a lot in the past (Camporeale, 2019). Most time-series forecasts (e.g., geomagnetic indices) are performed using artificial neural networks (e.g., Lethy et al. 2018). However, in the past few years, some convolutional architectures appeared to outperform recurrent neural networks in completing sequence modelling tasks (Bai et al. 2018). One of the issues behind forecasting near-Earth time-series using Lagrange 1 or BSN data lies in the propagation lead time between the two locations. Most forecasting methods solve this by prefixing which inputs from the past should be kept (e.g., McGranaghan et al., 2021): for instance, keeping only the values at $t_0-1$hr and $t_0-2$hr. Another idea is to use a variable propagation lead time based on the solar wind velocity (Wintoft et al., 2017). A TCN solves this problem by considering almost all past values and by finding, by itself, the relevant past values that will influence a given outcome. This paper is intended to be a proof of concept of the use of convolutional architectures to handle a variable lead time. It suggests that TCNs convincingly perform across a broad range of Space Weather time-series modelling.
16:45	Forecasting hazardous geomagnetically induced currents for Spanish critical infrastructures by using AI	Conde villatoro, D et al.	Oral
	Daniel Conde Villatoro[1], Florencia Luciana Castillo[2], Veronica Sanz González[1], Carmen García García[1], Bryan Zaldivar Montero[1], Jose Enrique García Navarro[1], Carlos Escobar Ibáñez[1]
	[1]Instituto de Física Corpuscular (IFIC); [2]Heidelberg University
	In the last decades, our society has become more interdependent and complex than ever before. Local impacts can cause global issues, as the pandemic clearly has shown, affecting the health of millions of human beings. It is also highly dependent on relevant technological structures, such as communications, transport, or power distribution networks, which can be very vulnerable to the effects of solar activity and their associated events, such as solar flares and coronal mass ejections, which may provoke disturbances, interruptions, and even long-term damage to these technical infrastructures, with drastic social, economic and even political impacts. However, these phenomena and their effects are not yet well understood, and their forecast is still in the early stages of development. In particular, geomagnetically induced currents (GICs) produced by solar storms can have catastrophic consequences on modern technology. This work, which uses a multidisciplinary approach, aims to deeply understand and develop an early warning system to evaluate the impact of violent solar storms on Spanish critical infrastructures. Specifically, we are developing an advanced machine-learning based predictive model of the impact of future solar storms on the ground. This model consists of two distinct stages. First, we are using as input real-time data from the solar wind space probe ACE (located at the L1 point in space) together with data from the geomagnetic observatories of Ebro, San Pablo de los Montes and San Fernando in Spain, and Coimbra in Portugal to develop a deep-learning model taking into account past conditions to predict the variation of the magnetic field on the Earth's surface at different locations in the Iberian Peninsula. Second, we will feed these local predictions of time-variation of the magnetic field into a physical model of the 3D Earth's geoelectrical structure to generate the geoelectrical fields that drive the GICs. Thus, the ultimate goal is to provide a real-time prediction of the GICs from extreme geomagnetic storms on the Spanish critical infrastructures.

Tuesday October 25, 17:00 - 18:00, Earth Hall

17:00	Decontamination of proton flux measurements in the radiation belts with machine learning	Bernoux, G et al.	Oral
	Guillerme Bernoux[1], Victor Le Couteur[2], Antoine Brunet[1]
	[1]ONERA / DPHY - Université de Toulouse - F31055 Toulouse; [2]Institut Supérieur de l'Aéronautique et de l'Espace, Université de Toulouse, 31400 Toulouse
	Some radiation monitors measuring charged particle fluxes (proton, electrons) in the radiation belts are subject to a contamination problem. These detectors can produce erroneous measurements because some populations are mistaken for others. This is the case of the Space Environment Monitor / Medium range energy proton detector (SEM/MEPED) instrument onboard the NOAA-14 satellite, that produces E > 2.5 MeV proton flux measurements contaminated by electrons, which are also monitored by the SEM instrument suite. This is a problem as such kind of data is useful for both space climate and weather studies. However, the data provided by the NOAA-15 satellite (which carries the SEM-2 suite) for the same energy ranges are much more reliable. Here we propose to exploit the overlap period of the two satellites to train a neural network to decontaminate NOAA-14 proton flux measurements using NOAA-15 measurements as reference. We show that, under certain conditions, our model is able to effectively decontaminate the measurements, but that some dedicated techniques are required to overcome the low volume of available overlapping data. This method could be extended to other contaminated datasets for which there is a reference dataset with an overlap period.
17:15	Convolutional Neural Networks for Automated ULF Wave Classification in Swarm Time Series	Antonopoulou , A et al.	Oral
	Alexandra Antonopoulou[1,2], Georgios Balasis[1], Constantinos Papadimitriou[1,2,3], Adamantia Zoe Boutsi[1,2], Athanasios Rontogiannis[4], Konstantinos Koutroumbas[1], Ioannis A. Daglis[2,5], Omiros Giannakis[1]
	[1]Institute for Astronomy, Astrophysics, Space Applications and Remote Sensing, National Observatory of Athens, Greece; [2]Department of Physics, National and Kapodistrian University of Athens, Greece; [3]Space Applications & Research Consultancy, SPARC G.P., Athens, Greece; [4]School of Electrical and Computer Engineering, National Technical University of Athens, Greece; [5]Hellenic Space Centre, Athens, Greece
	Ultra-low frequency (ULF) magnetospheric plasma waves play a key role in the dynamics of the Earth’s magnetosphere and, therefore, their importance in Space Weather studies is indisputable. Magnetic field measurements from recent multi-satellite missions (e.g. Cluster, THEMIS, Van Allen Probes and Swarm) are currently advancing our knowledge on the physics of ULF waves. In particular, Swarm satellites, one of the most successful missions for the study of the near-Earth electromagnetic environment, have contributed to the expansion of data availability in the topside ionosphere, stimulating much recent progress in this area. Coupled with the new successful developments in artificial intelligence (AI), we are now able to use more robust approaches devoted to automated ULF wave event identification and classification. The goal of this effort is to use a popular machine learning method, widely used in Earth Observation domain for classification of satellite images, to solve a Space Physics classification problem, namely to identify ULF wave events using magnetic field data from Swarm. We construct a Convolutional Neural Network (ConvNet) that takes as input the wavelet spectrum of the Earth’s magnetic field variations per track, as measured by Swarm, and whose building blocks consist of two alternating convolution and pooling layers, and one fully connected layer, aiming to classify ULF wave events within four different possible signal categories: 1) Pc3 wave events (i.e., frequency range 20-100 MHz), 2) background noise, 3) false positives, and 4) plasma instabilities. Our preliminary experiments show promising results, yielding successful identification of more than 97% accuracy. The same methodology can be easily applied to magnetometer data from other satellite missions and ground-based arrays.
17:30	Ensemble Learning for Accurate and Reliable Uncertainty Quantification	Camporeale, E et al.	Oral
	Enrico Camporeale
	University of Colorado & NOAA Space Weather Prediction Center
	In this presentation, we give an overview of the NASA/NSF funded “Ensemble Learning for Accurate and Reliable Uncertainty Quantification” SWQU project. The objective of the project is to create the algorithmic prototype that will combine a small number of high-fidelity (low error, but computationally expensive) runs from physics-based models with a large number of (possibly) less accurate but faster runs from machine learning models. The goal is to obtain a more accurate overall prediction than any individual model, and an estimation of the associated uncertainties. We showcase application examples ranging from geomagnetic index predictions to solar wind and ground based magnetic field forecasting. The core engine of the probabilistic predictions is the Accurate and Reliable Uncertainty Estimate (ACCRUE) model, that is briefly discussed. ACCRUE is able to estimate uncertainties associated to deterministic predictions, by solving a deep learning semi-supervised regression problem
17:45	Short-term forecasting of Total Electron Content in South America	Perez bello, D et al.	Oral
	Dinibel Perez[1,2], Ma Paula Natali [1,2], Amalia Meza [1,2], Luciano Mendoza [1,2]
	[1] Laboratorio de Meteorología Espacial, Atmósfera Terrestre, Geodesia, Geodinámica, diseño de Instrumental y Astrometría (MAGGIA), Facultad de Ciencias Astronómicas y Geofísicas (FCAG), Universidad Nacional de La Plata (UNLP), La Plata, Argentina, [2] Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET), Buenos Aires, Argentina
	The term space weather describes the variations in the environment between the Sun and Earth. This complex interaction is one of the principal sources of the ionosphere’s dynamics. The behavior of this part of the atmosphere represents one of the main concerns of modern society because it could threaten the proper working of modern technological infrastructure, e.g., communication networks, power grids, global navigation systems, oil pipelines. Therefore, monitoring and forecasting the Earth’s upper atmosphere behavior is a subject of increasing attention \cite{chen2021prediction}. Several instruments provide information to describe the state and dynamics of the ionospheric plasma. These include ionosondes, incoherent radars, Faraday rotation, GNSS, space-borne sensors, among others. In this group, the GNSS technique is the only one that provides a 24x365 global coverage through the freely accessible global GNSS tracking infrastructure that contributes to the International GNSS Service (IGS). From these observations, it is possible to obtain the vertical total electron content (vTEC) and generate vTEC maps with a high time and space resolution. MAGGIA laboratory produces near-real-time regional vTEC maps \cite{mendoza2019maggia} every 15 minutes incorporating data from approximately 80 GNSS satellites tracked by more than 200 ground stations. In this work, we present a vTEC forecast model for quiet and disturbed days for 2019. The input data are the regional vTEC maps produced by MAGGIA laboratory in a grid of 0.5º by 0.5º in longitude and latitude. The forecasting model consists of several steps. First, an interpolation algorithm is adapted to repair missing data. Second, an autoregressive nonlinear neural network (NAR-NN) is applied to compute vTEC multiple steps ahead in each grid point. The forecast was implemented with different horizons in advance using 1-hour sampling. Finally, interpolation techniques are studied to optimize the processing time without losing the resolution of MAGGIA maps.
18:00	A method to choose a mother wavelet for feature detection of VLF signals for Machine learning	Shivali, V et al.	Oral
	Shivali Verma, Sonendra Gupta
	Oriental College of Technology, Bhopal -462021,India.[1.2]
	In present age Artificial Intelligence is demanding technique in the space physics. It is possible through the learning of artificial neurons that make machine learning possible. Therefore, for the learning of artificial neurons input must be given in precise form. In this work we have identified the mother wavelet which help to find out the best mother for feature extraction of Very low frequency (VLF) emissions frequency range 2 kHz to 20 kHz. These emissions are play imperative role for probing of ionosphere and magnetosphere. There are three main VLF signals whistler, chorus and hiss. They all have specific characteristics with respective time and frequency. Wavelet analysis is being a widespread time-frequency analysis method. An effort to propose a more competent analysis of the VLF signal, the solicitation of the wavelet analysis is deliberated. The encounter in wavelet analysis is choosing the best mother wavelet for evaluating the signals, as many mother wavelets applied on the signal may yield diverse results. In this work we have used DEMETER satellite obtained VLF signals for the wavelet analysis. For “optimal" choice of the mother wavelet for such signals we have applied energy based quantitative approach at different frequency level of VLF signals which are obtained by Discrete Wavelet Transform (DWT). Results show that Haar and Bior3.5 are the optimum mother wavelet for detection of feature for VLF signal. Further VLF feature are extracted by identified wavelets and used to developed ANN VLF classifier. The output of the classifier is finally use to study the VLF emission emitted during the geomagnetic storm occur on August 24, 2005 when DETMETER ply over the India region.

Posters

1	Surrogate Modeling for Faster Space Weather Prediction	Baeke, H et al.	Poster
	Hanne Baeke [1], Jorge Amaya [2], Giovanni Lapenta [3]
	[1], [2], [3] Centre for Mathematical Plasma Astrophysics, KU Leuven, Heverlee, Belgium.
	Plasma simulations require often a lot of computing time and resources. Therefore, I am looking for a faster alternative. I am developing a machine learning surrogate model in order to mimic plasma simulations, requiring less computing time without sacrificing precision. The use of surrogate models for space physics is still in its infancy, but shows already promising results in other fields. The poster will present the current developments, some first results and the future plans of my research. The surrogate model is based on the one introduced by Sanchez et al. (2020), connected to DeepMind. This paper uses Graph Neural Networks (GNN) to create a surrogate model of SPH simulations of falling water, sand and goop and their interactions with each other and a surrounding box. I want to extend this method to plasma simulations, which have as additional difficulty that the interactions are not only governed by collisions, but also by the electric and magnetic field, as well as the particles’ charges. A fluid particle-in-cell code, called Slurm (Olshevsky et al. 2019), is used to generate the simulations for different plasma conditions. The output of these simulations serves as training data for the Graph Neural Networks surrogate model. The machine learning method will first be applied on small scale plasma phenomena, like the Kelvin-Helmholtz instability or magnetic reconnection. Later on, the method will be expanded to large scale plasma physics and will be applied to the magnetosphere around Mercury and the Moon. This will aid fast space weather forecasting around these bodies, useful for future manned missions to the Moon for example. If the method turns out to be successful, it could also be applied to other environments, modeling for example the space weather conditions between the Sun and Earth. Olshevsky V., Bacchini F., Poedts S., Lapenta G. Slurm: Fluid particle-in-cell code for plasma modeling. Comput. Phys. Comm., 235 (2019), p.16. Sanchez-Gonzalez A., Godwin J., Pfaff T., Ying R., Leskovec J., Battaglia P.W. Learning to simulate complex physics with graph networks. Proceedings of the 37th International Conference on Machine Learning, 784 (2020), p.8459-8468.
5	Can Machine Learning solve the „Bz Problem“ in Interplanetary Coronal Mass Ejections?	Reiss, M et al.	Poster
	Martin A. Reiss[1], Christian Möstl[1], Rachel Bailey[2], Hannah Rüdisser[3], Ute Amerstorfer[1], Tanja Amerstorfer[1], Andreas Weiss[1], Jürgen Hinterreiter[1], and Andreas Windisch[3]
	[1]Space Research Institute, Austrian Academy of Sciences, Graz, Austria; [2]Zentralanstalt für Meteorologie und Geodynamik, Vienna, Austria; [3]Know-Center GmbH, Graz, Austria.
	Predicting the Bz magnetic field embedded in interplanetary coronal mass ejections (ICMEs), also called the Bz problem, is a critical challenge in space weather research and forecasting. We tackle this problem with a new approach by taking upstream in situ measurements of the ICME sheath region and the first few hours of the magnetic obstacle to predict the downstream Bz component. To do so, we trained machine learning algorithms on 348 ICMEs (extracted from the open-source ICMECATv2.0 catalog) observed by the Wind, STEREO-A, and STEREO-B satellites to predict the minimum value of Bz. The predictive tool was built to mimic a real-time scenario, where the ICMEs sweep over the spacecraft, which allows us to continually provide updates and improved predictions of Bz as time passes and more of the CME structure is observed. The final model, which is based on random forests, can predict the minimum value of Bz with a reasonable level of agreement compared to observations. In this presentation, we will discuss the main challenges we face in using a data-driven machine learning application to solve the Bz problem, and outline the lessons learned and future strategies for predicting and potentially mitigating the effects of ICMEs arriving at Earth.
7	A Comparative Study on New ML Approaches for F10.7 Time Series Forecasting	Marcucci, A et al.	Poster
	Adriana Marcucci[1], Giovanna Jerse[2], Isacco Zinna[3], Marco Molinaro[2], Mauro Messerotti[2]
	[1]University of Trieste, Department of Physics; [2]Astronomical Observatory of Trieste, INAF; [3]University of Trieste, Department of Mathematics and Geo-science
	Solar radio flux at 2.8 GHz (F10.7) is a commonly used proxy for solar activity and one of the main drivers of the Operational Space Weather. Specifically it can seriously affect both the ionospheric Total Electron Content (TEC) - whose variations are sources of GNSS signals quality degradation - and the thermospheric density, which perturbs the orbits of objects in low Earth orbit. It is very crucial to estimate in advance the solar effects on most applications and services but the analysis and prediction of solar radio emissions are challenging tasks as the underpinning physical processes are typically nonlinear, non-stationary and chaotic. Various F10.7 index forecasting methods have been published but in this work we present an original approach for short and medium-term forecasting up to 45 days, based on the hybrid combination of decomposition techniques with Long Short Time Memory (LSTM) neural networks. In particular the Fast Iterative Filtering (FIF) method is employed to decompose the original F10.7 timeseries and the Multi Head Attention module is introduced. Results of comparative analysis with other models available in the literature highlight strengths and weaknesses of each model under different conditions, such as solar activity level, lead-time, and required computational resources to train them.
8	Using machine learning to predict the timing, magnitude, and impact of solar flares.	Edward-inatimi, N et al.	Poster
	Nathaniel Edward-Inatimi [1], Ciaran Beggan [2]
	[1] University of Edinburgh, s1739049@ed.ac.uk, [2] British Geological Survey, Edinburgh, UK, ciar@bgs.ac.uk
	Solar flares are highly energetic and stochastic events originating on the Sun's surface. They are a key component of space weather hazard which can impact the Earth in a number of ways from high frequency (HF) radio blackouts, to geomagnetically induced currents (GIC) through geomagnetic storms. Using a timeseries of the well-known SDO HMI flaring indicator, called ‘R’, a convolutional neural network (CNN) was developed and trained to predict solar flare occurrence within a defined flaring window of between 3 and 48 hours. The training data for the CNN was created using R parameter time series accessed for all flaring and non-flaring active regions (ARs) listed within the GOES XRS reports from 2013 to 2017. A data set consisting of 220 M-class (and stronger) flaring samples with an equal number of non-flaring samples was created. The most effective model was found to have a 12-hour flaring window, yielding a True Skill Statistic (TSS) score of 0.71 with an accuracy of 85% (compared with a baseline accuracy of 52%). Whilst not a definitive result, the model achieves comparable TSS scores to previous machine learning efforts with a relatively limited training data set. There is scope to expand the CNN to account for flare class and to fit this research into the wider context of flare prediction and space weather impact mitigation.
9	Towards explanation of airglow variation by ML techniques	Varga, M et al.	Poster
	Matej Varga[1],Simon Mackovjak[1],Peter Butka[2],Viera Maslej-Krešňáková[2],Samuel Amrich[3],Adrián Kundrát[2]
	[1]Institute of Experimental Physics Slovak Academy of Sciences, [2]Faculty of Electrical Engineering and Informatics Technical University of Košice, [3]Faculty of Mathematics and Physics Charles University
	The Earth's upper atmosphere acts as an interface between processes in space and on Earth. It is a very dynamic environment continuously influenced by solar radiation and space weather from above and by atmospheric weather and electrical discharges from below. To describe processes in this interface environment is a challenging task that requires consideration of a very wide range of phenomena. To overcome this challenge, we have developed a data-driven approach based on state-of-the-art machine learning techniques to model important thermosphere-ionosphere characteristic - airglow radiation (Mackovjak et al., 2021a). We realized that this data-driven approach might be even more precise if we involve additional information that are not commonly available. For this reason, we have developed SCSS-net - one of the most precise model for solar corona structures segmentation based on deep neural networks (Mackovjak et al., 2021b) that allow automatic characterization of solar activity with high temporal and spatial resolution. We have also developed a deep learning approach for automatic detection of tweek atmospherics (Maslej-Krešňáková et al., 2021) and a system for autonomous detection of TLE's (Amrich et al., 2021), generated by lightning strikes, that provide additional information about the state of the lower ionosphere. The main points of all these results will be presented and new ongoing challenges will be discussed. - Mackovjak et al.: 2021a, JGR Space Phys., 126, 3, https://doi.org/10.1029/2020JA028991 - Mackovjak et al.: 2021b, Mon. Not. R. Astron. Soc., 508, 3, https://doi.org/10.1093/mnras/stab2536 - Maslej-Krešňáková et al.: 2021, Earth Space Sci., 8, 11, https://doi.org/10.1029/2021EA002007 - Amrich et al.: 2021, J. Instrum, 16, T12016, https://doi.org/10.1088/1748-0221/16/12/T12016
10	Applications of artificial intelligence in studies of space weather.	Asimopolos, L et al.	Poster
	Laurentiu Asimopolos [1], Natalia-Silvia Asimopolos [1],
	[1] Geological Institute of Romania
	s