S O I - PROPOSAL INVESTIGATION TITLE: Neural Network Studies of SOHO Magnetograms LEAD INVESTIGATOR: Henrik Lundstedt, Lund Observatory, Box 43, S-221 00 Lund, e.mail: hlundstedt@solar.stanford.edu, WWW: http://nastol.astro.lu.se/~henrik/spacew1.html, phone: +46 46 222 7294, fax: +46 46 222 46 14 TEAM MEMBERS: Peter Wintoft (Lund), J.-G. Wu (Lund), Hans Gleisner (Lund), Todd Hoeksema (CSSA/Stanford), Xuepu Zhao (CSSA/Stanford) and Ludwik Liszka (Swedish Institute of Space Physics). SOI COORDINATOR: Todd Hoeksema SSSC PROGRAMMER: - ABSTRACT/ TECHNICAL SUMMARY: We intend to study, the large-scale organization of the solar magnetic flux and time series of magnetograms with the use of neural networks. The results from the studies will then be incorporated in a hybrid intelligent system which predicts the space weather from solar magnetic field observations. Along the study we also hope to obtain improved and deepened knowledge about fundamental problems within solar physics. The study will take advantage of the unique quality of the solar magnetic field observations made by SOHO. The magnetograms are characterized by both high spatial (full-disk (4") and 11'x11' field (1.25")) and high temporal (every 96 minutes and every minute during 8-hour campaigns) resolution and by not being influenced by the degrading effects of the Earth's atmosphere. These factors are of vital importance for our studies. INVESTIGATION PLAN: The first study aims at a better understanding of how the large-scale magnetic flux is organized. The study will tell us about conditions deep below solar surface. This obtained new knowledge will then be compared with helioseisomological results. The origin of the organization is another fundamental question, that has been addressed by e.g. Stenflo (1991), Hoeksema (1991) and that also our study will address. Finally another related question is when and where active longitudes emerge, a question which is of great importance for predictions of the space weather? We will use both supervised and unsupervised neural networks, which are capable of learning and finding organization and clustering of fluxes of different sizes. Unsupervised neural networks have been used by Snel and Lundstedt (1993) for clustering of solar wind parameters and of solar activity structures in WSO synoptic charts by Wintoft and Lundstedt (1993). Since coronal phenomena, such as coronal mass ejections (CMEs) and coronal holes, are associated with the large-scale structures and changes of these structures, we also hope to learn more about these coronal phenomena and predictions of them. The second study aims at a better understanding of the solar activity. Questions that will be addressed are: Is the solar dynamic system chaotic (Mundt et al.,1991) or not? What dimension has the system? What characterizes the solar activity in terms of changes of flux sizes, distribution and emerging rates? Are there new periodicities we haven't observed yet? Before applying the neural networks for time series analysis of magnetograms we intend to preprocess the data according to new methods discovered by Liszka (1995a,b). These methods better show the information in the data. Our neural network approach makes it then possible to study an enormous information space. New solar physics is therefore likely to be the outcome of the study. The amount of data however demands dimension reducing neural networks or the use of neural chips. Both methods will be applied. The goal is then finally to use the newly reached knowledge from the two studies to develop a hybrid intelligent system, which predicts the space weather (Lundstedt et al., 1995). The developed hybrid intelligent system is foreseen to be of big practical importance. The hybrid intelligent system will consist of neural networks, fuzzy expert systems and could also include MHD models. Neural networks will predict CMEs and solar activity from magnetograms. MHD-models (Zhao and Hoeksema, 1993) have been used to successfully calculate the coronal magnetic fields and the heliospheric current sheet. The output from the neural networks and calculations could then be input to other neural networks and fuzzy expert systems which will finally predict solar wind parameters (Wintoft and Lundstedt 1993; Zhao and Hoeksema, 1995) and geomagnetic activity (Wu and Lundstedt, 1995). Accurate predictions of the space weather (e.g. geomagnetic storms) 1-3 days in advance are very difficult to accomplish (Joselyn, 1995). However, we think the use of all knowledge obtained from these studies of SOHO magnetograms, combined in a hybrid intelligent system, could vastly improve the accuracy. REFERENCES: Hoeksema, J.T.:1991, Global Solar Magnetic Fields, J. Geomagn. and Geoelectr., 43. Joselyn, J.:1995, Geomagnetic activity forecasting: The state of art, Rev. Geophys., AGU, 33, 3. Liszka, L.:1995a, Reconstruction of Equidistant Time Series Using Neural Networks, Scientific Report, Swedish Institute of Space Physics. Liszka, L.:1995b, Analysis of Multivariate Time Series, Scientific Report, Swedish Institute of Space Physics. Lundstedt, H., Wintoft, P., Wu, J.-G. and Gleisner, H.:1995, AI Methods and Space Weather Forecasting, in proceedings of Artificial Intelligence Knowledge Based Systems for Space, 5th Workshop 10-11 October 1995 ESTEC, Noordwijk, The Netherlands. Mundt, M. D., Maguire, W.B. and Chase, R.P. P.:1991, Chaos in the Sunspot Cycle: Analysis and Prediction, J. Geophys. Res. Vol. 96, 1705-1716. Snel, R. and Lundstedt,H.:1993, Self Organizing maps of Solar Wind Structures, in proceedings of the International Workshop on Artificial Intelligence Applications in Solar-Terrestrial Physics, eds., J. Joselyn, H. Lundstedt and J. Trolinger, Lund, Sweden 22-24 September 1993. Stenflo, J.:1991, Weak Solar Magnetic Fields, in D.S. Spicer (ed.), Electromechanical Coupling of the Solar Atmosphere, Proc. QSL Workshop, Capri, Italy, May 27-31, 1991. Wintoft, P. and Lundstedt, H.:1993, Geomagnetic Activity and Large-Scale Solar Magnetic Field Structures Studied With Two Neural Network Paradigms, in proceedings of the International Workshop on Artificial Intelligence Applications in Solar-Terrestrial Physics, eds., J. Joselyn, H. Lundstedt and J. Trolinger, Lund, Sweden 22-24 September 1993. Wintoft, P. and Lundstedt, H.:1995, IMF polarity prediction from WSO magnetograms from WSO magnetograms using radial basis neural network, presented at the IUGG XXI General Assembly meeting in Boulder, July 2-14, 1995, p B140. Wu, J.-G. and Lundstedt, H.:1995, Predictions of geomagnetic storms from solar wind data using Elman recurrent neural networks, submitted to Geophys., Res. Letters. Zhao X. P. and Hoeksema, J.T.: 1993, A Coronal Magnetic Field Model With Horizontal Volume and Current Sheet, Solar Phys., 143, 41. Zhao X. P. and Hoeksema, J.T.:1995, Prediction of the interplanetary magnetic field strength, J. Geophys. Res., 100, 19-33.