Dieser Inhalt ist zur Zeit nur in englischer Sprache verfügbar. Machine learning (ML) is a rapidly developing field of research that has advanced various domains such as fraud detection, web search results, credit scoring, automation, email spam filtering, and many others. Unlike the traditional methods efficient mostly at linear problems, the ML methods and algorithms reached their success because of being effective at nonlinear problems with multiple input variables. Another important aspect is that ML relies on the abundance of data. In our research, we apply machine learning methods and algorithms to space weather related problems in order to develop new data-driven models of different physical phenomena in space and to enhance existing physics-based models. These problems and processes are commonly highly nonlinear and also plentiful of data due to the rising awareness of space weather and an increasing number of scientific missions studying those phenomena.
In our group, we use ML algorithms to develop predictive models of electron density in the plasmasphere [Zhelavskaya et al., 2016; Zhelavskaya et al., 2017] and Kp index. In both of these projects, feedforward neural networks are employed. They are a very powerful tool for finding nonlinear multivariate mappings from input to output parameters. In the case of electron density, the inputs to the model are the time history of geomagnetic indices (AE, Kp, Sym-H, and F10.7) [Zhelavskaya et al., 2017], and in the case of Kp index, the inputs are the time history of solar wind and IMF parameters. The figure shown here is an example of electron density reconstruction for 26 June 2001 22:27 UT. On the left you can see the image of the He+ particles distribution in the plasmasphere taken by the EUV instrument on board of the IMAGE satellite mapped to the equatorial plane. White dots indicate the plasmapause location identified manually from the image. On the right, the reconstruction of electron density using a neural network for the same time is shown. Color indicates the decimal logarithm of electron density, and the black and gray section of the colorbar indicates the approximate position of the plasmapause. The actual plasmapause location taken from the IMAGE EUV data is also overplotted with the white dots. As it can be seen from the figure, the density reconstruction obtained using the neural network matches the reality quite well. This time period is out of sample (i.e., was not included in the training).
In this project, we will utilize the developed models of plasma density and Kp index to enhance the forecast of the radiation belts. Specifically, we will incorporate the models of electron density and Kp forecast into the Versatile Electron Radiation Belt (VERB) code developed by our group that will allow for a more precise prediction of the radiation belt state.
Zhelavskaya, I. S., M. Spasojevic, Y. Y. Shprits, and W. S. Kurth (2016), Automated determination of electron density from electric field measurements on the Van Allen Probes spacecraft, J. Geophys. Res. Space Physics, 121, 4611–4625, doi:10.1002/2015JA022132.
Zhelavskaya I. S., Y. Y. Shprits, and M. Spasojevic (2017), Empirical modeling of the plasmasphere dynamics using neural networks, J. Geophys. Res., 122, doi:10.1002/2017JA024406.
"Empirical modeling of the plasmasphere dynamics using neural networks",
Space Weather: A Multi-Disciplinary Approach Workshop at the Lorentz Center, September 25-29, 2017, Leiden, the Netherlands.
"Deriving electron density from electric field measurements on the Van Allen Probes spacecraft and building a global dynamic model of plasma density using neural networks"
IAPSO-IAMAS-IAGA 2017, August 27 - September 1, 2017, Cape Town, South Africa.
"Global dynamic evolution of the cold plasma inferred with neural networks", DGG 2017, March 27-30, 2017, Potsdam, Germany.