## Session 13 - System Science: Application to space weather analysis, modelling and forecasting

Richard Boynton (Univ of Sheffield); Homayon Aryan (Goddard Space Flight Center); George Balasis (IAASARS, NOA); Enrico Camporeale (CWI)
Friday 01/12, 9:45 - 13:00
Delvaux

KEYWORDS - System science, machine learning, data assimilation, information theory, signal processing

The construction of accurate dynamical models is fundamental to forecasting the many aspects of space weather. The process of deducing models for forecasting traditionally involved breaking the system into component parts and applying the laws of physics to each part to build up a description of that system. However, for complex space weather systems, we do not have enough knowledge about some of the processes involved to build an accurate model solely from first principles. Alternatively, complex systems science based methods have been developed to deduce dynamical models from input-output data. The techniques developed from system science, such as system identification, machine learning, data assimilation, information theory, signal processing, among others, are applicable to any system that has large amounts of data availability. With the increasing amount of space weather data, we are able to make use of these tools and techniques to analyse, model and forecast the complex systems of space weather. As such, this session is for contributions that employ these state of the art tools that have been developed in system science.

Poster Viewing
From Thursday morning to Friday noon

Talks
Friday December 1, 09:45 - 11:00, Delvaux
Friday December 1, 11:45 - 13:00, Delvaux

### Talks : Time schedule

Friday December 1, 09:45 - 11:00, Delvaux
 09:45 Understanding performance of neural network models for short-term predictions applied to geomagnetic indices Wintoft, P et al. Invited Oral Peter Wintoft, Magnus Wik Swedish Institute of Space Physics Many processes in the solar-terrestrial chain relevant for space weather are nonlinear with temporal dynamics (memory). Different nonlinear system identification approaches, such as support vector machines, neural network models, and NARMAX models, have successfully been applied to the analysis and prediction in space weather. They perform very well compared to numerical MHD models (Rast\"{a}tter et al., 2013) and in the prediction mode they are computationally light-weight which means that large number of events can be analysed and they can be implemented for real-time operation. In this work we specifically focus on neural networks. A great challenge is the collection and preprocessing of data from which the mathematical relations can be extracted and tested. This includes studying the distribution of data and identifying the tails (extremes), a task that becomes complex in multi-dimensional space (e.g. solar wind plasma and magnetic fields), and especially also when the temporal evolution needs to be considered. For example, a point in an one-dimensional case may lie well inside from the tail, but when embedded in state-space the point can become extreme. The extremes are important to study because they are often related to extreme effects but also because they are the most difficult from a modelling perspective. To some degree tackle the difficulties at the extremes we use an ensemble of models from which the median prediction is used. The data set size and distribution put limits on the accuracy and operational range of the models. The models performs best in the bulk of the distribution and can not extrapolate outside the observed distribution and it is important to understand when this happens and possibly look at simpler alternative models. We will describe the latest developments along that described above applied to the short-term predictions of geomagnetic indices \emph{Kp}, \emph{Dst}, \emph{AE}, and local geomagnetic $dB/dt$. We will also compare with various simple relations for the driving phase of the storm such as $V B_{\bot} \sin^4 \theta/2$ and $R_\mathrm{quick}$ (Borovsky and Birn, 2013). To be able to perform a more detailed analysis of the storm events we apply a time series wavelet analysis in order to determine range, length, and phases of individual storms. Finally, we describe the verification carried out be able to asses the performance. This work has in part been supported by ESA SSA Space Weather ESC contract No 4000113185/15/D/MRP and European Union's Horizon 2020 grant agreement No 637302 (PROGRESS). 10:10 Automatically worked stages for estimation of the level of expected radiation hazards from SEP events Dorman, L et al. Oral Lev Dorman[1,2] and Lev Pustil’nik[1] [1]Israel Cosmic Ray and Space Weather Center, affiliated to Tel Aviv University, Israel Space Agency, and Shamir Research Institute; [2]Institute of Terrestrial Magnetism, Ionosphere and Radiowave Propagation RAN of N.V. Pushkov (IZMIRAN), Moscow, RU-142190, Russia In the last years thanks to formation of NMDB became possible to have on-line through Internet one-minute cosmic ray (CR) data from many neutron monitors and muon telescopes (in high energy region) as well as from several spacecrafts (in very low energy region). To avoid damage of electronics and negative effects for people health is necessary on-line forecast expected fluency of energetic particles and radiation hazards. It was shown by myself and colleagues in many original research papers that this possible to do by using the first 20-30 minutes of CR data on the basis of coupling functions, spectrographic method, and by solving inverse problem, and then calculate expected results on radiation hazards for many hours of great Solar Energetic Particle (SEP) event. Usually it takes a lot of time (at least, several months). But for really protection of satellites, aircrafts, and people from dangerous radiation hazards all this must be made automatically, including the formation of corresponding alerts on the expected level of radiation hazard for different objects. We describe several automatically worked stages and obtain corresponding algorithms. The first stage works continue, collecting from Internet all available one minute data on CR variations (corrected on meteorological and geomagnetic effects). The second stage also works continue according to automatically working program "SEP-Start" - supposed, developed and checked in the Israel Cosmic Ray and Space Weather Center. Using of this program on many CR stations and on satellites allowed to determine automatically the beginning of SEP event (it can be different at different stations caused to anisotropy at beginning of SEP). If the second stage gives positive result, starts to work automatically the third stage according to program "SEP-Coupling" – using method of coupling functions and spectrographic method for transformation obtained at different altitudes and cutoff rigidities data on CR intensity variations to the space and calculation CR energy spectrum and angle distribution out of the Earth’s atmosphere and magnetosphere, directly in the interplanetary space near the Earth. After obtaining results by third stage starts to work automatically the fourth stage according to program "SEP-Inverse Problem", and it is determined source function, time of ejection SEP into interplanetary space, and diffusion coefficient of propagation in dependence of SEP energy and distance from the Sun. After obtaining results by fourth stage starts to work automatically the fifth stage according to program "SEP-Direct Problem", and it is determined by found at fourth stage parameters the time variations of primary SEP in dependence of particles energy in interplanetary space near the Earth for many hours ahead, up to few days (on the basis of only 20-30 minutes of SEP beginning). On the basis of information, obtained in the fifth stage, it is easy to calculate by known coupling functions and cutoff rigidities expected time variations of SEP intensity in SPACE- and AIR-CRAFTS at different trajectories, and compare the beginning part with available observations and estimate the quality of forecasting (sixth stage, program “SEP-Forecasting”). If the forecasted radiation hazard is expected dangerous for different objects, will be immediately send corresponding Alerts (seventh stage, program “SEP-Alerts”). With time by obtaining new data, forecasting Alerts became more and more exactly. We prepared all algorithms to realize this program automatically. We consider also how to extend the program for more complicated cases, including anisotropic SEP events. Keywords: great SEP events, radiation hazards, automatically forecasting, algorithms 10:30 Empirical modeling of the plasmasphere dynamics using neural networks Zhelavskaya, I et al. Oral Irina Zhelavskaya[1,2],Yuri Shprits[1,2,3],Maria Spasojevic[4] [1]Helmholtz Centre Potsdam, GFZ German Research Centre For Geosciences, Potsdam, Germany,[2]University of Potsdam, Potsdam, Germany,[3]University of California, Los Angeles, Los Angeles, CA, USA,[4]Hansen Experimental Physics Laboratory, Stanford University, CA, USA We propose a new empirical model for reconstructing the global dynamics of the cold plasma density distribution based only on solar wind data and geomagnetic indices. Utilizing the density database obtained using the NURD (Neural-network-based Upper hybrid Resonance Determination) algorithm for the period of October 1, 2012 - July 1, 2016, in conjunction with solar wind data and geomagnetic indices, we develop a neural network model that is capable of globally reconstructing the dynamics of the cold plasma density distribution for 2 ≤ L ≤ 6 and all local times. We validate and test the model by measuring its performance on independent datasets withheld from the training set and by comparing the model predicted global evolution with global images of He+ distribution in the Earth’s plasmasphere from the IMAGE Extreme UltraViolet (EUV) instrument. We identify the parameters that best quantify the plasmasphere dynamics by training and comparing multiple neural networks with different combinations of input parameters (geomagnetic indices, solar wind data, and different durations of their time history). We demonstrate results of both local and global plasma density reconstruction. This study illustrates how global dynamics can be reconstructed from local in-situ observations by using machine learning techniques. 10:45 Variation of geomagnetic responses to the solar wind input: Half-year increase during declining phases and 2009 specialty Yamauchi, M et al. Oral M. Yamauchi[1] and B. Olsthoorn[2] [1]Swedish Institute of Space Physics (IRF), Kiruna; [2]Master course student, Department of Physics, Stockholm University Variations of the Sun-Earth coupling efficiency (AL, AU, ASY-D, and SYM-H indices for the same level of solar wind electromagnetic energy input determined by Akasofu's energy coupling function "epsilon") since 1981 were examined using NASA/OMNI 5-min data. The seasonal variation was removed by averaging the indices for given ranges of the epsilon values over every three months around equinoxes and solstices. (1) For small to moderate epsilon, the AL and AU responses to the same epsilon shortly increased beyond the fluctuation level for about half a year for both equinoxes and solstices during 1983, 1994, 2003, and late 2015, all during early declining phase of solar cycles, although the timing does not necessarily the same as the peak solar wind energy input; (2) Except these singular periods, this Sun-Earth coupling efficiency for small to moderate epsilon continuously decreased over the past three decades until 2009, and then started to recover afterword. (3) The short increase during the singular periods and long-term trend with 2009 low are also found in the mid-altitude ASY-D index, but are not clear in SYM-H; The long term change in the Sun-Earth coupling efficiency raises a possibility that it can be related to the strength of the solar cycle. If this is true, the strength of the solar cycle 25 will somewhat recover from current solar cycle 24.

Friday December 1, 11:45 - 13:00, Delvaux