Session 13 - Machine Leaning and statistical inference techniques applied to space weather

Giovanni Lapenta (KULeuven), Enrico Camporeale
Thursday 21/11, 11:15-12:30 & 17:15-18:30
Rogier

The science of 'making predictions' has been historically based on statistical inference (e.g., frequentist, Bayesian, information criterion-based) and, more recently, on machine learning techniques.
Entire disciplines, such as system identification, data assimilation, information theory, deep learning and uncertainty quantification, have proliferated in the attempt to improve our ability to extract information from data and build predictive models.
Each of these disciplines has been studied and developed in contexts typically unrelated to Space Weather (e.g., quantum mechanics, financial forecasting, astronomy, etc.), yet present powerful new opportunities for our community. Coupled with massively expanded data availability and sophisticated means to analyze voluminous and complex information, the timing is ripe for the Space Weather community to embrace new innovative methodologies.

This session is devoted to contributions to Space Weather specification and prediction that use innovative, multidisciplinary, and, perhaps, unconventional approaches.

Talks
Thursday November 21, 11:15 - 12:30, Rogier
Thursday November 21, 17:15 - 18:30, Rogier

Click here to toggle abstract display in the schedule

Talks : Time schedule

Thursday November 21, 11:15 - 12:30, Rogier

11:15	Super-Resolving Magnetograms covering 40 years of space weather events	Shneider, C et al.	Oral
	Carl Shneider[1], Mandar Chandorkar[1], Enrico Camporeale[1,2].
	[1]CWI; [2]CIRES
	We apply deep learning architectures with the aim of converting magnetograms from a source survey to a target survey while preserving the features and systematics of the target survey. As a first step, we test the validity of a deep learning approach based on a single instrument conversion between different resolutions in order to utilize perfect alignment and identical systematics and identify the conditions under which the conversion breaks down. We perform this analysis on magnetograms taken by the Helioseismic and Magnetic Imager (SDO/HMI). We then apply the same approach to upscale and cross-calibrate magnetograms obtained by the Michelson Doppler Imager (MDI/SOHO), as well as magnetograms taken by the Global Oscillation Network Group (GONG) to the resolution of SDO/HMI.
11:30	A machine learning approach for automated ULF wave recognition	Balasis, G et al.	Oral
	Georgios Balasis[1], Sigiava Aminalragia-Giamini[1], Constantinos Papadimitriou[1], Ioannis A. Daglis[2], Anastasios Anastasiadis[1], Roger Haagmans[3]
	[1]National Observatory of Athens, [2]2National and Kapodistrian University of Athens, [3]European Space Agency
	Machine learning techniques have been successfully introduced in the fields of Space Physics and Space Weather, yielding highly promising results in modelling and predicting many disparate aspects of the geospace environment. Magnetospheric ultra-low frequency (ULF) waves can have a strong impact on the dynamics of charged particles in the radiation belts, which can affect satellite operation. Here, we employ a method based on Fuzzy Artificial Neural Networks in order to detect ULF waves in the time series of the magnetic field measurements on board the low-Earth orbit (LEO) CHAMP and Swarm satellite missions. The outputs of the method are validated against a previously established, wavelet-based, spectral analysis tool, that was designed to perform the same task, and show encouragingly high scores in the detection and correct classification of these signals.
11:45	What is the intrinsic dimensionality of the OMNI data? A dimensionality reduction study	Teunissen, J et al.	Oral
	Jannis Teunissen[1,2], Romain Dupuis[2], Carl Shneider[1], Enrico Camporeale[1,3]
	[1] CWI, Amsterdam, The Netherlands, [2] KU Leuven, Belgium, [3] University of Colorado Boulder, USA
	We apply dimensionality reduction to the low-resolution OMNI data set, which contains hourly averages of about 50 variables from the 1960s up to today. These variables for example describe near-Earth solar wind conditions, plasma parameters, and indices such as the Dst, Kp and F10.7 index. We compare the performance of different dimensionality reduction techniques, such as principal component analysis (PCA), auto-encoders and novel approaches such as UMAP, by studying how much of the variance in the original data can be explained from a smaller set of variables. Furthermore, we study whether dimensionality reduction for the OMNI data set is affected by the source of the solar wind (e.g. ejecta or coronal holes). Using an existing classification of the solar wind, we apply the reduction methods to subsets corresponding to different source regions. Finally, we investigate to what extent the performance of dimensionality reduction can be improved by taking the history of the 'remaining' variables into account, in other words by considering multiple hourly averages as input for the reduction. This contribution has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 776262 (AIDA, www.aida-space.eu )
12:00	Forecasting solar wind properties using dimensionality reduction and Self-Organizing Maps	Amaya, J et al.	Oral
	Jorge Amaya[1], Romain Dupuis[1], Jannis Teunissen[1][2], Giovanni Lapenta[1]
	[1] CmPA, Mathematics Department, KU Leuven, Belgium, [2] CWI, Amsterdam, The Netherlands
	The OMNI2 database contains information about the solar wind, including dozens of in-situ plasma properties and geomagnetic indices measured on Earth. Each entry in this database is a single point in a multi-dimensional feature space. Some of these features are more or less inter-correlated. We simplify the database using dimensionality reduction techniques to eliminate features that are highly correlated. We also apply Principal Component Analysis (PCA) to maximize the variability of the different axis of the simplified solar wind hyper-space. We then use machine learning to perform forecasts. Supervised machine learning techniques have been used in the past to forecast geomagnetic indices, with a special focus on the prediction of the Dst. Here we present a different approach: we use Self-Organizing Maps (SOM) to catalog and forecast solar wind properties. SOMs are an unsupervised machine learning technique used to map complex multi-dimensional points into a simple 2D matrix that retains topological information of the feature hyper-space. We use SOMs to analyze the simplified solar wind database described above. The 2D map generated is a representation of the geometrical distance between the solar wind properties at different times: similar solar wind properties will activate the same group of nodes in the map. Neighboring nodes in the map are also close together in the feature hyper-space. SOMs allow us to: catalog the properties into a small group of map nodes, evaluate the influence of individual properties on the activation of map nodes, and evaluate how empirical classification models developed in the past correlate with automatic methods. In addition, SOMs can be used to forecast future solar wind properties. We will show how once the model is trained, new solar wind data can activate map nodes, giving a good indication of the future plasma properties. This contribution has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 776262 (AIDA)
12:15	Assessing the predictability of the geomagnetic activity with information theoretical tools	Bernoux, G et al.	Oral
	Guillerme Bernoux[1], Antoine Brunet[1], Miho Janvier[2], Eric Buchlin[2]
	[1] ONERA / DPHY, Université de Toulouse, F-31055 Toulouse – France, [2] Institut d’Astrophysique Spatiale, CNRS, Univ. Paris-Sud, Université Paris-Saclay 91405 Orsay CEDEX – France
	Over recent years, several predictive models of terrestrial geomagnetic indices based on solar wind parameters at 1AU have been developed. The work carried out has shown strong limitations on the prediction horizon achievable by these methods. Indeed, the use of measurements made by satellites at the Lagrange L1 point does not allow a reliable prediction of geomagnetic activity beyond a 6-hour window. To extend this horizon to a window of up to a few days, we plan to use Sun images directly as inputs to Neural Networks for the longer-term prediction of the state of the magnetosphere. This study consists in analyzing the transport of information, in the sense of information theory, between solar wind parameters and geomagnetic indices so as to determine the most suitable data set for prediction purposes. Therefore we compute time-lagged multivariate metrics such as the so-called Mutual Information and Transfer Entropy while using data sets (including geomagnetic indices) from the OMNIWEB database, with the aim of extending this research to solar imaging made available by the Multi Experiment Data & Operation Center (MEDOC). We also evaluate the impact of different gap-filling techniques, such as Singular Spectral Analysis, on the accuracy of our results.

Thursday November 21, 17:15 - 18:30, Rogier

17:15	A gray-box model for a probabilistic estimate of regional ground magnetic perturbations: Enhancing the NOAA operational Geospace model with machine learning	Camporeale, E et al.	Oral
	Enrico Camporeale[1,2,3], Michele D. Cash[2], Howard J. Singer[2], Christopher C Balch[2], Zhenguang Huang[4], Gabor Toth[4]
	[1] CIRES, University of Colorado, Boulder, [2] NOAA/Space Weather Prediction Center, [3] CWI, Amsterdam, [4] University of Michigan
	Geomagnetically induced currents (GIC) represent one of the most significant effects caused by space weather events on the ground. They are generated when a geomagnetic storm produces a sudden perturbation in the Earth’s magnetic field that, through Faraday’s law, induces an electric field along conductors such as power grids and pipelines. GIC are also one of the most difficult phenomena to predict from a physical standpoint, because of their large variability in time, and the intricacies of the Earth’s magnetic field and its nonlinear coupling with the surrounding space plasma. In this work we present a model that forecasts the maximum value of magnetic perturbation (specifically, the horizontal component of dB/dt) over 20 minutes interval. The model builds on the results of the physics-based Geospace model, developed at University of Michigan, that runs operationally at NOAA’s Space Weather Prediction Center. By using the simulation output as an input for a machine learning classification method we are able to predict the probability that dB/dt will exceed a given threshold, on a specific location.
17:30	Multivariate Timeseries Analysis for Solar Flare and Eruption Forecasting: the Unexploited Potential and its Blending with Machine Learning	Georgoulis, M et al.	Oral
	M. K. Georgoulis[1,2], R. A. Angryk[3], P. C. Martens[1], B. Aydin[3], A. Ahmadzadeh[3], R. Ma[3]
	[1] Department of Physics and Astronomy, Georgia State University, Atlanta, GA 30303, USA, [2] RCAAM of the Academy of Athens, 11527 Athens, Greece, [3] Computer Science Department, Georgia State University, Atlanta, GA 30303, USA
	We often overlook the undisputed fact that solar flares and eruptions are outcomes of an evolutionary course in solar source regions, be them active regions or the quiet Sun. In essence, this is the archetypical storage and release process of magnetic energy and helicity that are achieved within timescales of days or weeks in the low solar atmosphere but are released explosively within minutes or hours. Nonetheless, within today’s plethora of forecasting methods and techniques, only a slim minority exploits the temporal evolution of forecast parameters or properties, in most cases at a minimal level. Recent results have solidified that operational forecasts gain value when time series are involved in the process. We therefore describe recent and ongoing efforts at the Georgia State University (GSU) Astroinformatics Cluster to exploit detailed, high-cadence and multivariate time series in a big data environment, namely the Solar Dynamics Observatory (SDO) HMI Active Region Patch (SHARP) properties and more, to advance forecasting of eruptive solar activity. Besides conventional time series descriptors, techniques such as Dynamic Time Warping in the framework of machine learning methodologies may offer ways to meaningfully improve machine cognition reflected on higher-level performance verification skill scores. Publicly accessible multivariate time series appropriate for this task are continuously created, both in the US (namely, the GSU solar flare benchmark dataset) and in Europe (namely, the FLARECAST property database and others). Given the existing data, we foresee a proliferation of such methodologies in the future, pushing the limits of conventional, point-in-time forecasting capabilities.
17:45	Geomagnetic Kp index forecast using historical values and real-time observations	Shprits, Y et al.	Oral
	Yuri Shprits[1], Ruggero Vasile[1], Irina Zhelavskaya[1]
	[1] GFZ-Potsdam, German Research Center for Geocsciences
	Current algorithms for the real-time prediction of the $Kp$ index use a combination of models empirically driven by solar wind measurements at the L1 Lagrange point and historical values of the index. In this study, we explore the limitations of this approach, examining the forecast for short and long lead times using measurements at L1 and $Kp$ time series as input to artificial neural networks. We explore the relative efficiency of the solar wind-based predictions, predictions based on recurrence, and based on persistence. Our modeling results show that for short-term forecasts of approximately half a day, the addition of the historical values of $Kp$ to the measured solar wind values provides a barely noticeable improvement. For a longer-term forecast of more than two days, predictions can be made using recurrence only, while solar wind measurements provide very little improvement for a forecast with long horizon times. We also examine predictions for disturbed and quiet geomagnetic activity conditions. Our results show that the paucity of historical measurements of the solar wind for high $Kp$ results in a lower accuracy of predictions during disturbed conditions. Rebalancing of input data can help tailor the predictions for more disturbed conditions.
18:00	Supervised machine learning for flare prediction: the impact of features and of the training set generation process on the forecasting performances	Piana, M et al.	Oral
	Cristina Campi[1], Federico Benvenuto[2], Anna Maria Massone[2,3], Manolis Georgoulis[4,5], D Shaun Bloomfield[6], Michele Piana[2,3]
	[1] Dipartimento di Matematica, Università di Padova, [2] Dipartimento di Matematica, Università di Genova, [3] CNR - SPIN Genova, [4]Department of Physics and Astronomy, Georgia State University, [5] RCAAM of the Academy of Athens, [6] Department of Mathematics, Physics and Electrical Engineering, Northumbria University
	Solar flare prediction may now benefit of very accurate and extensive measurements of solar full disk line-of-sight and vector magnetograms, the ones provided by SDO/HMI, and of a rich set of active region properties, the ones produced by the pattern recognition algorithms developed within the Horizon 2020 FLARECAST project. This talk exploits this unprecedented wealth of information to train different supervised machine learning algorithms and to show that: 1) the performances of these methods significantly depend on the biases introduced in the process of generation of training sets; 2) the information hidden in the magnetic properties of the active regions are intrinsically redundant and a significantly small fraction of features extracted from these data provides an actual contribution to forecasting; 3) when appropriately trained, these algorithms produce prediction skill scores significantly smaller than one, thus confirming the substantially stochastic nature of solar flares.
18:15	Progress and issues predicting the Dst index using Long Short-Term Memory neural networks.	Laperre, B et al.	Oral
	Brecht Laperre[1], Jorge Amaya[1], Giovanni Lapenta[1]
	[1]KU Leuven
	We use Artificial Neural Networks (ANN) to forecast the Dst index several hours in advance, using as input solar wind observations at 1AU. Inspired by state-of-the-art models used in automatic translation and image captioning, we use Long-Short Term Memory (LSTM) ANNs to treat the prediction of the Dst as a sequence-to-sequence analysis, using a multitude of different architectures to achieve our goal. As input, we use a window of time that is situated 6 hours before the to-be predicted value, containing measurements from the solar wind. The goal of the LSTM model is to ”translate” this time series into the Dst series of the upcoming 6 hours. In our work, we also focus on the problem of neural network learning the persistence model. This work has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 776262 (AIDA, www.aida-space.eu)

Posters

1	Automatic Generation of Daily Space Environment Forecast Text Based on Natural Language Generation	Zou, Y et al.	p-Poster
	Yenan Zou, Jingjing Wang, Yanxia Cai, Siqing Liu
	National Space Science Center, Chinese Academy of Sciences
	The Space Environment Prediction Center of the Chinese Academy of Sciences has been providing daily space environment forecasts to the public and space agencies since its establishment in 1992. At present, the space environment summary and space environment forecast text in daily space environment forecast are generated semi-automatically by forecasters' analysis of observation data and model prediction results. We present in this paper an application which automatically generates textual daily space environment forecast, using the observed data released by NOAA. This solution combines data analysis methods and strategies for linguistic description of data together with a natural language generation system in an innovative way. On the one hand, the application extracts spatial environment element information from observation data and encodes it into intermediate descriptions using linguistic variables. These descriptions are later translated into spatial environment summary text by natural language generation system. On the other hand, the application integrates spatial environment model forecast and artificial experience forecast results to generate spatial environment forecast text by natural language generation system. In the presentation we will show more details about the solution.
2	The Rate of Change of the Surface Magnetic Field in the UK: Sources and Forecasting	Smith, A et al.	p-Poster
	A. W. Smith[1], I. J. Rae[1], C. Forsyth[1], M. P. Freeman[2]
	[1]MSSL/UCL, [2]British Antarctic Survey
	Rapid changes in the surface geomagnetic field can induce potentially damaging currents in conductors on the ground; this is a critical consideration for the operation of power networks and pipelines. Several physical drivers of such field variability exist, including solar wind pressure pulses, geomagnetic storms and substorms. In this work we investigate the physical sources of the largest rates of change of the horizontal magnetic field (R) recorded by three UK based ground stations. We then investigate possible methods of forecasting intervals of large R using prior observations of the solar wind and geomagnetic indices. Firstly, we classify the physical causes of the largest rates of change measured by UK magnetometers using multivariate classification/clustering techniques. The process of classification further enables the robust selection of relevant parameter inputs for the purposes of forecasting. The forecasting problem is then considered as a multivariate timeseries prediction: given the preceding solar wind and geomagnetic conditions, what are the expected future values of R? Models including LSTM (Long Short-Term Memory) neural networks are evaluated and compared to persistence and ARIMA (AutoRegressive Integrated Moving Average) models. Finally, it is useful to predict whether a threshold value of R will be exceeded during a future interval, given the observed solar wind and geomagnetic conditions. To this end ensemble models such as decision trees are evaluated and compared to climatological predictions.
3	Using LSTM neural networks to forecast geomagnetic storms	Peeperkorn, J et al.	p-Poster
	Jari Peeperkorn,Romain Dupuis,Giovanni Lapenta
	KU Leuven,KU Leuven,KU Leuven
	Introduction: The goal of this work is to forecast geomagnetic storms 3-6 hours before they happen. Geomagnetic storms are disturbances in Earth's magnetic field caused by solar winds or coronal mass ejections, traveling from the Sun towards the Earth. These disturbances are created due to an injection of plasma into the magnetosphere and the subsequent increase in the currents located in the ionosphere. These geomagnetic storms are being closely monitored, due to there possible effects of human technology. For example geomagnetically induced currents in power systems can cause black-outs and changes in the density of the ionosphere can cause mishaps in navigation and communication systems. Geomagnetic storms are measured with different indices, the most prominently used being the Disturbance Storm Time index (or Dst). The Dst measures the decrease of the magnetic field at low latitude, caused by an increase in the ring current. Different models have been used over the years to forecast the Dst and in recent years the use of machine learning, more specifically artificial neural networks, has made some promising results. \\ Method: The Dst is forecasted with the help of two different sets of data, consisting of past values of different input features. These features are selected with a correlation analysis. Both sets contain the Dst itself, the magnitude of the interplanetary magnetic field and its z-component. The first also contains solar wind parameters velocity, temperature and proton density. The second set is complemented with other geomagnetic indices. The model used is a long short-term memory neural network, a variant of a recurrent artificial neural network designed to perform better on long term dependencies. A hyperparameter grid search was done to fine tune the model, before training and testing it. \\ Results: The results obtained by using the metrics root mean square error (RMSE) and Pearson's correlation coefficient (CC) look promising and are similar to the best results found in other literature. However by visually checking the forecasts on 4 independent geomagnetic storms it can be seen that the models do not always provide a sufficiently reliable forecast that can be used for mitigation (especially 4-6 hours in advance). Based on this distinction it was opted to also use a binary classification that divides the data into two categories: storm hour and no-storm hour. This was done with two different thresholds: one using the classification of a moderate storm and one using the classification of an intense storm. Both of these thresholds show a relatively high precision but a bad recall rate. This means that the model's forecast does not capture a lot of the storm hours correctly. Discussion: The LSTM model shows potential to use in geomagnetic storms forecasting, purely based on the results of the RMSE and CC. These results also compare well with other results from literature. However based on the bad recall rates and the visual check of 4 storm examples it can be concluded that the model is not reliable. This indicates as well that RMSE and CC, used a lot in literature, do not adequately measure the true performance of the model.
4	Analyzing big data from space missions and massively parallel simulations within the Horizon 2020 Project AIDA	Lapenta, G et al.	p-Poster
	Giovanni Lapenta [1], AIDA Consortium (www.aida-space.eu)
	[1] KULeuven - Belgium , [2] UNIVERSITA' DELLA CALABRIA - Italy, [3] Università di Pisa - Italy, [4] CWI, CENTRUM WISKUNDE & INFORMATICA- The Netherlands, [5] SYSTHMATA YPOLOGISTIKIS ORASHS IRIDA LAB- Greece, [6] CINECA CONSORZIO INTERUNIVERSITARI- Italy, [7] SPACE CONSULTING INTERNATIONAL LLC - United States [8] CENTRE NATIONAL DE LA RECHERCHE SCIENTIFIQUE CNRS - France
	AIDA (www.aida-space.eu) is a new H2020 project funded by the European Commission to develop new techniques to analyze large data sets from space missions and high performance computing simulations using statistical and machine learning techniques. Missions and simulations are progressively generating more and more and larger and larger data sets. Space science and space weather forecasting require the access to as much data input as possible to drive physics-based and heuristic models. Progress has increased the size and resolution of the observations while simulations are increasing their size producing ever growing data sets. The space community is undergoing two transitions that are common to other areas of science. First, the use of the python free-access language is increasing exponentially with new packages being made available to the community to replace some of the tasks previously delegated to mission-specific programs often written in proprietary (and very expensive) languages such as IDL. Python is making data accessible to everybody without the hurdle of purchasing IDL. But the more important advantage of transitioning to python is that by accessing and manipulating space data via python, a plethora of tools developed in a myriad field becomes available to the space community. Foremost among them machine learning and artificial intelligence. Second, machine learning (ML), deep learning (DL) and artificial intelligence (AI) are taking over the world by storm. Space is not an exception with many new projects and activities starting on applying various ML, DL and AI methods. Success stories already abound and great opportunities lay ahead for the analysis of the large data set typical of space with these methods. AIDA, Artificial Intelligence Data Analysis, aims at speeding up the progress of space science along the two paths described above. AIDA develops the AIDApy package to integrate many of the functions so far done via IDL, filling the gaps of other community based efforts in neighboring fields to provide the functions specifically needed to space scientists without replicating functions already available in other packages. The AIDA effort needs feedback from the community to achieve this task, The most ambitious task of AIDA is to then use the AIDApy to bring to the space community the most advanced ML, DL and AI tools developed in the information technology business making these advanced accessible to a community not trained in computer science. AIDA chose among the various ML frameworks, PyTorch, developed by Facebook. We will outline the reasons for this choice and present some examples of the first applications based on AIDApy. This work has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 776262 (AIDA, www.aida-space.eu)
5	NARMAX approach to the development of spatiotemporal models for space weather forecast.	Balikhin, M et al.	p-Poster
	Michael A. Balikhin [1], Richard J. Boynton [1]
	[1] SSL, ACSE, The University of Sheffield
	Approach based on Nonlinear AutoregRessive Moving Average models with eXogenous inputs (NARMAX models) has been successfully applied for the development of space weather forecasting tools for various space weather parameters such as geomagnetic indices, daily fluxes of energetic electrons at GEO and others. In these examples the evolution of a particular space weather parameter can treated as the output of SISO (Single Input -Single Output) or MISO (Multi Input -Single Output) system. In this presentation the recent advances in application of NARMAX to the forecast of distributed geospace parameters (parameters that vary both in time and space) are reviewed. Forecasting models for magnetospheric emissions and for spatiotemporal evolution of fluxes of energetic electrons in the magnetosphere will be used to illustrate two complementary NARMAX based approaches to the prediction of space weather parameters that evolve both in space and time.
6	Flare Prediction using Deep Learning with multiple wavelength SDO data	Koukras, A et al.	p-Poster
	Alexandros Koukras [1][2], Laurent Dolla [2], Benoit Frénay [3]
	[1] KU Leuven, [2] Royal Observatory of Belgium, [3] UNamur
	Goal: Our goal is to utilize the state of the art of deep learning (DL) in image recognition algorithms to predict the flaring activity of the Sun. In combination with the prediction we aim to identify and examine the most prominent physical features that are indications of a possible flare. Scope: There have been many attempts to predict flaring activity of the Sun with different methods and data. But almost all of them use as input the calculated features extracted from line-of-sight magnetograms . Only, recently there seems to be an interest in the automatic detection of features and the use of a ML subfield, called Deep Learning. We attempt to make a probabilistic prediction of a specific class flare, using a Convolutional Neural Network (CNN). The basic role of the CNN is to automatically detect features from the images, instead of hand-picking different features for input. Using SDO observations we create a training dataset of flaring and non-flaring active regions, which is used to train the CNN. The performance of the prediction is estimated using multiple forecast verification metrics (Sensitivity, Accuracy, False-Alarm ratio, Heidke skill score, True Skill Statistics and more). The novelty of this work is based in the additional use of EUV images in multiple channels for the training of the CNN instead of magnetogram data only. This is useful because a number of studies have shown that there are valuable information, for the prediction of flares, in the higher layers of the Sun’s atmosphere. Another novelty of this work could be considered the treatment of the CNN as a source of information and not as a black box. To accomplish that we look back at the detected features that are activated when there is a classification of a certain type of flare. This is possible through the visualization of the feature map of the CNN, which is a common technique in image recognition algorithms.
7	Using dynamical networks to characterize and quantify the evolving spatio-temporal ground pattern of magnetic disturbance seen by 100+ ground based magnetometers with SuperMAG	Chapman, S et al.	p-Poster
	Sandra Chapman [1], Lauren Orr [1], Jesper Gjerloev [2,3]
	[1] Centre for Fusion, Space and Astrophysics, Physics Dept., University of Warwick, Coventry CV4 7AL, UK [2] Johns Hopkins University Applied Physics Laboratory, Laurel, Maryland, USA [3] Department of Physics and Technology, University of Bergen, Bergen, Norway
	Networks provide a generic methodology to characterize spatio-temporal pattern in large datasets of multi-point observations. Networks are now a common analysis tool in societal data where it is clear whether two nodes are connected to each other or not. In observations from real physical systems, a ‘connection’, meaning significant cross-correlation or coherence between the timeseries seen at two observation points, is more subtle to establish. We have developed methodology to construct dynamical directed networks of the SuperMAG 100+ magnetometers for the first time. If the canonical cross-correlation (CCC) between vector magnetic field perturbations observed at two magnetometer stations exceeds a threshold, they form a network connection. The time lag at which CCC is maximal determines the direction of propagation or expansion of the structure captured by the network connection. If spatial correlation reflects ionospheric current patterns, network properties can test different models for the evolving current system during geomagnetic storms and substorms. Importantly, once the network is constructed, one can quantify it with a few parameters. One such parameter is the normalized degree (average number of connections each station or group of stations has to all others in the network) which can be split into different latitude bands and regions w.r.t. the auroral oval and, during a substorm, the substorm current wedge. Such parameters make possible quantitative statistical comparisons of hundreds of substorms and storms that capture the full spatial distribution of activity, without relying on gridding or infilling of data. In the first applications of this network methodology we have obtained the timings of the propagation of the ground magnetic signal of a northward or southward turning of the interplanetary magnetic field [1] and have established the characteristic evolution of the ground disturbance pattern seen in 86 isolated substorm events [2]. [1] J. Dods, S. C. Chapman, J. W. Gjerloev, Characterising the Ionospheric Current Pattern Response to Southward and Northward IMF Turnings with Dynamical SuperMAG Correlation Networks, JGR, 122, doi:10.1002/2016JA023686. (2017) [2] L. Orr, S. C. Chapman, J. Gjerloev, Directed network of substorms using SuperMAG ground-based magnetometer data, GRL submitted (2019)
8	Flare Prediction using Deep Learning with multiple wavelength SDO data	Koukras, A et al.	p-Poster
	Alexandros Koukras [1][2], Laurent Dolla [2], Benoit Frénay [3]
	[1] KU Leuven, [2] Royal Observatory of Belgium, [3] UNamur
	Goal: Our goal is to utilize the state of the art of deep learning (DL) in image recognition algorithms to predict the flaring activity of the Sun. In combination with the prediction we aim to identify and examine the most prominent physical features that are indications of a possible flare. Scope: There have been many attempts to predict flaring activity of the Sun with different methods and data. But almost all of them use as input the calculated features extracted from line-of-sight magnetograms . Only, recently there seems to be an interest in the automatic detection of features and the use of a ML subfield, called Deep Learning. We attempt to make a probabilistic prediction of a specific class flare, using a Convolutional Neural Network (CNN). The basic role of the CNN is to automatically detect features from the images, instead of hand-picking different features for input. Using SDO observations we create a training dataset of flaring and non-flaring active regions, which is used to train the CNN. The performance of the prediction is estimated using multiple forecast verification metrics (Sensitivity, Accuracy, False-Alarm ratio, Heidke skill score, True Skill Statistics and more). The novelty of this work is based in the additional use of EUV images in multiple channels for the training of the CNN instead of magnetogram data only. This is useful because a number of studies have shown that there are valuable information, for the prediction of flares, in the higher layers of the Sun’s atmosphere. Another novelty of this work could be considered the treatment of the CNN as a source of information and not as a black box. To accomplish that we look back at the detected features that are activated when there is a classification of a certain type of flare. This is possible through the visualization of the feature map of the CNN, which is a common technique in image recognition algorithms.
9	Complex systems perspectives pertaining to the research of space weather	Balasis, G et al.	p-Poster
	Georgios Balasis[1], Reik V. Donner[2], Jakob Runge[3]
	[1]National Observatory of Athens, [2]Magdeburg-Stendal University of Applied Sciences, [3]Institute of Data Science, German Aerospace Center (DLR)
	Learning from successful applications of methods originating in statistical mechanics or information theory in one scientific field (e.g. atmospheric physics or weather) can provide important insights or conceptual ideas for other areas (e.g. in the space sciences or beyond) or even stimulate new research questions and approaches. For instance, quantification and attribution of dynamical complexity in output time series of nonlinear dynamical systems is a key challenge across scientific disciplines. Especially in the field of space physics, an early and accurate detection of characteristic dissimilarity between normal and abnormal states (e.g. pre-storm activity vs. magnetic storms) has the potential to vastly improve space weather diagnosis and, consequently, the mitigation of space weather hazards. Information theory techniques have great potentials to identify previously unrecognized precursory structures and, thus, to contribute to a better understanding of the evolution of geomagnetic field perturbations along with extreme space weather phenomena. In this presentation, we utilize a variety of complementary modern complex systems based approaches to obtain an entirely novel view on nonlinear magnetospheric variability. We first show successful applications of nonlinear measures based on the analysis of recurrences of previous states to studying the Dst index along with characteristic variables of the solar wind around intense magnetic storms. Moreover, the time-dependent coupling between the solar wind and the magnetosphere along with the relationship between magnetic storms and magnetospheric substorms is of paramount importance for space weather processes. However, the storm/substorm relationship is one of the most controversial aspects of magnetospheric dynamics. In order to further disentangle this relationship and the role of relevant solar wind variables as drivers and mediators, multivariate causality measures employing the concept of graphical models constitute one particularly promising tool. Toward this goal we then highlight the great potential of combining a causal discovery algorithm with a multivariate and lag-specific extension of transfer entropy for tackling contemporary research questions in magnetospheric physics, such as the storm-substorm relationship.
10	Prediction of extreme flaring events using machine learning methods	Piana, M et al.	p-Poster
	Federico Benvenuto[1], Cristina Campi[2], Anna Maria Massone[1,3], Michele Piana[1,3]
	[1] Dipartimento di Matematica, Università di Genova, [2] Dipartimento di Matematica, Università di Padova, [3] CNR - SPIN Genova
	Thanks to the effort produced within the framework of the Horizon 2020 FLARECAST project, several supervised and unsupervised machine learning algorithms are now at disposal for solar flare predictions. We employed such notable computational corpus to show that machine learning allows a rather reliable and detailed forecasting of the sequence of flaring events associated to a specific solar storm. Specifically, this potentiality is illustrated in the case of the September 2017 eruptions by using feature vectors corresponding to active region properties extracted from SDO/HMI magnetograms.
11	Identification of magnetic reconnection regions in PIC simulations with machine learning	Dupuis, R et al.	p-Poster
	Romain Dupuis[1], Jorge Amaya[1], Giovanni Lapenta[1]
	KU Leuven, Belgium
	Magnetic reconnection is a fundamental process for many plasma phenomena. It can convert the stored magnetic energy into thermal and non thermal energy by restructuring the magnetic field topology. For instance, this process can be responsible of coronal mass ejection (CME), very important in space weather. Several detection algorithms and signatures of reconnection regions[1], such as the agyrotropy [2], have been developed in the literature, usually using various field quantities. We propose here to apply machine learning techniques to particle distributions in order to identify reconnection regions. Density estimations methods, in particular Gaussian mixture model, are coupled to metrics from information theory and applied to PIC simulations. Gaussian mixtures models assumes the distribution function is built as a combination of several Gaussian distributions with unknown parameters. The latter are determined by the algorithm using the data points from the distribution. The Gaussian mixtures is able to automatically detect distribution functions with complex shapes, such as beams and can clearly identify shapes very specific to reconnection. The results can then be compared to other simulations or data provided by the Magnetospheric Multiscale (MMS) mision [3]. This contribution has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 776262 (AIDA, www.aida-space.eu) [1] Cazzola, et al. (2016). On the electron agyrotropy during rapid asymmetric magnetic island coalescence in presence of a guide field. Geophysical Research Letters, 43(15), 7840-7849. [2] Swisdak, M. (2016). Quantifying gyrotropy in magnetic reconnection. Geophysical Research Letters, 43(1), 43-49. [3] Burch, J. L., Torbert, R. B., Phan, T. D., Chen, L. J., Moore, T. E., Ergun, R. E., ... & Wang, S. (2016). Electron-scale measurements of magnetic reconnection in space. Science, 352(6290)
12	Classification of Magnetosheath Jets using Neural Networks, Solar Wind Observations and High-resolution IMF Measurements.	Raptis, S et al.	p-Poster
	Savvas Raptis[1], Sigiava Aminalragia-Giamini[2], Tomas Karlsson[1], Per Arne Lennart Lindqvist[1]
	[1]Space and Plasma Physics, School of Electrical Engineering and Computer Science, KTH Royal Institute of Technology, Sweden, [2]Space Applications & Research Consultancy (SPARC), Greece
	Magnetosheath jets are enhancements of dynamic pressure resulting from solar wind interaction with the bow shock and indicating a locally increased plasma flow. Jets are believed to be a key element in the coupling between the solar wind and the magnetosphere while also being associated with other physical phenomena such as magnetic reconnection, radiations belts, ionospheric flow enhancements and throat auroral features. All these phenomena are directly connected to space weather field, thus making jets a key research component. In this work, we use a dataset that includes several thousands of magnetosheath jets that have been classified into four classes. The first two main categories are jets found in the Quasi-parallel magnetosheath $(θ_Bn<45°)$ and those found in the Quasi-perpendicular $(θ_Bn>45°)$, with $θ_Bn$ being the angle between the Interplanetary Magnetic Field (IMF) and the bow shock’s normal vector. Two more categories are based on different transitions between the mentioned cases. “Boundary” jets are found when we have a switch from quasi-parallel bowshock to quasi-perpendicular or vice versa, and “Encapsulated” jets are jets holding quasi-parallel characteristics while the surrounding plasma before and after the jet is of quasi-perpendicular nature. This initial dataset has been derived by using in-situ measurements of various plasma moment quantities and magnetosheath magnetic field as measured by the Magnetospheric Multiscale (MMS) mission during 09/2015 – 05/2019. We then use solar wind data, measured outside of the magnetosheath, in L1, in order to predict the four (4) classes of the jets that were later observed inside the magnetosheath region by MMS. The predictive classification is done with deep Neural Networks (NNs) and several different inputs including several solar wind particle moments, electric field, and IMF values. Preliminary results already support the initial classification scheme of the magnetosheath jets. More importantly, they show that even in the absence of crucial information, such as the angle of the IMF, the use of machine learning methods allow a connection between the solar wind particle population before and after its complex interaction with Earth’s bow shock as measured by different missions.
13	Leveraging the Mathematics of Shape for Machine Learning Prediction of Solar Magnetic Eruptions	Berger, T et al.	p-Poster
	Thomas Berger[1], Varad Deshmukh[1,2], Elizabeth Bradley[2], James Meiss[3]
	[1]University of Colorado at Boulder, Space Weather TREC, [2]University of Colorado at Boulder Department of Computer Science, [3]University of Colorado at Boulder Department of Applied Mathematics
	Current operational forecasts of solar eruptions are primarily made by human experts using qualitative shape-based classification systems and historical data about flaring frequencies. In the past decade, there has been a great deal of interest in crafting machine-learning (ML) flare-prediction methods using algorithms that extract underlying patterns from a training set---e.g., a set of solar magnetogram images, each characterized by features derived from the magnetic field and labeled as to whether it was an eruption precursor. These patterns, captured by various methods (neural nets, support vector machines, etc.), can then be used to classify new images. A major challenge with any ML method is the featurization of the data: pre-processing the raw images to extract higher-level properties, such as characteristics of the magnetic field, that can streamline the training and use of these methods. It is key to choose features that are informative, from the standpoint of the task at hand. To date, the majority of ML-based solar eruption methods have used physics-based magnetic and electric field features such as the total unsigned magnetic flux, the gradients of the fields, the vertical current density, etc. In this paper, we extend the relevant feature set to include characteristics of the magnetic field that are based purely on the geometry and topology of 2D magnetogram images and show that this improves the prediction accuracy of a neural-net based flare-prediction method.
15	Convolutional Neural Networks for Automated Detection of ULF Waves in Swarm Time Series	Antonopoulou, A et al.	p-Poster
	A. Antonopoulou, C. Papadimitriou, A. Z. Boutsi, K. Koutroumbas, A. Rontogiannis, O. Giannakis and G. Balasis
	National Observatory of Athens
	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 mission 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 deep learning method in order to classify ULF wave events using magnetic field data from Swarm. We construct a Convolutional Neural Network (CNN) that takes as input the wavelet spectra of the Earth’s magnetic field variations per track, as measured by each one of the three Swarm satellites, and whose building blocks consist of two convolution layers, two pooling layers and a standard NN layer, aiming to classify ULF wave events in four different categories: 1) Pc3 wave events (i.e., frequency range 20-100 MHz), 2) non-events, 3) false positives, and 4) plasma instabilities. Our primary experiments show promising results, yielding successful identification of more than 95% accuracy. We are currently working on producing larger training/test datasets, by analyzing Swarm data from the mid-2014 onwards, when the final constellation was formed, aiming to construct a dataset comprising of more than 50000 wavelet image inputs for our network.