Science Success Story

« Back

XSEDE Allocation Facilitates Expanded Solar Wind Predictions

Novel model uses machine learning to better predict geomagnetic storms

By Kim Bruch, San Diego Supercomputer Center (SDSC)

NASA image of the Sun-Earth system transformed into the style of the painting Udnie by Francis Picabia (1913) using the neural style transfer tool online. Design by Andong Hu, CWI, Amsterdam, and it's the cover of Chandorkar's PhD thesis.

While space weather can produce dancing lights here on Earth, such as the beautiful Aurora borealis and Aurora australis that sometimes streak across the northernmost and southernmost skies, geomagnetic storms can cause severe damage to power grids, satellites, and many other electrical systems.

Heliophysicist Bala Poduval has dedicated her space weather research to predicting the solar wind conditions that cause these storms, with a recent highlight being the validation of a machine learning model that she and her colleagues developed for solar wind prediction in which simulated data using the Comet supercomputer at the San Diego Supercomputer Center (SDSC) played a pivotal role.

These comparisons of predicted and actual hourly solar wind forecasts, for a period spanning approximately one month (11/16/16-12/14/16), were generated using data computed on Comet at the San Diego Supercomputer Center. Credit: Mandar Chandorkar, Cyril Furtlehner, Bala Poduval, Enrico Camporeale, and Michele Sebag.

"To mitigate the adverse effects of space weather, it is necessary to forecast these events with sufficient lead time so that appropriate safety precautions can be taken," said Poduval, a research scientist with the University of New Hampshire. "One of the important pieces of information we needed to improve our prediction model was the ambient solar wind velocity near the Earth."

Upon receiving an allocation from the National Science Foundation's Extreme Science and Engineering Discovery Environment (XSEDE), Poduval and her international colleagues utilized Comet to validate a model that predicts time-lagged effects of the solar wind and the dependence on the wind's velocity. To accomplish this, they used a machine learning technique called Dynamic Time Lag Regression (DTLR) and validated their work by predicting the solar wind arrival near the Earth's orbit from physical parameters of the Sun as measured from the ground and space.

"After being referred to XSEDE by a colleague, I was able to submit a proposal for supercomputing time and received a start-up allocation in only a couple of weeks," said Bala Poduval, a research scientist with the University of New Hampshire. "I would like to thank SDSC's User Services Group's Mahidhar Tatineni as he helped me with creating a script for optimizing the computational time."

Bala Poduval, a heliophysicist at the University of New Hampshire, recently used SDSC's Comet to better understand solar wind prediction. Her research focuses on the prediction of solar wind conditions using coronal extrapolation models. She is also exploring methods of artificial intelligence for substantially improving the accuracy of prediction. Credit: Bala Poduval.

"We used the changes in the solar magnetic field, in terms of the magnetic flux tube expansion factor, as well as other characteristics of the sun such as the number of sunspots, to determine the speed of the solar wind that can be detected at Earth after a lag of ‘dynamic time'," explained Poduval. "The validation of DTLR and the solar wind forecasting would not have been possible in any reasonable time frame without XSEDE high performance computing resources since the computation of magnetic flux expansion factor for a period sufficient to train a neural network would take months on a local machine as compared to just a few weeks with Comet."

The code for this research was written in Interactive Data Language (IDL) and takes at least five hours to run a single model to run on a personal computer. Scaling up Poduval's work required over one thousand simulations, which would have taken over 5000 hours on a personal computer. "We not only saved computational time thanks to XSEDE, but were also able to optimize our code thanks to SDSC staff," said Poduval. "Further, we were able to complete two papers thanks to XSEDE and I am really thankful for this amazing service available for our academic research community."

Poduval and her colleagues' work was recently presented at the International Conference on Learning Representations, which was held in a virtual format due to the COVID-19 pandemic. Poduval is now exploring methods of artificial intelligence for substantially improving the accuracy of prediction, and with a recent National Science Foundation award, she is developing a neural network model for predicting solar energetic particle events.

This research was funded by a grant from the Centrum Wiskunde and Informatica, Amsterdam. Computational work on Comet was allocated by the National Science Foundation's XSEDE (TG-ATM170013) Initiative.

At a Glance

  • SDSC's Comet was used to validate a model that predicts time-lagged effects of a cause (of solar origin), which in this case was the velocity of solar wind.
  • Dynamic Time Lag Regression (DTLR) allowed for validation by predicting the solar wind arrival near the Earth's orbit from physical parameters of the Sun as measured from the ground and space.
  • XSEDE played a critical role in this research by providing an allocation to conduct the study in a timely manner.