CME Society

https://colab.research.google.com/drive/1cB8Py_yhkPDgjibmB0mLfUJQzVZXl6vw?usp=sharing

https://drive.google.com/file/d/1EEA-Jj2FgZ70gJoSGGEyrlrk4GFvu676/view?usp=sharing

Website Link: here

Slideshow Link: here

Detailed Project Video (1:45): here

How we addressed this challenge:

Our team created Neutron, our neural network, which uses scikit-learn's MLPRegressor Model to predict solar wind speed, solar wind intensity, and solar wind temperature. This model uses filtered data from DSCOVR and WIND that was processed with custom scripts. Our method of filtering data creates much cleaner training data, which results in a better model than if we used data directly from the satellites.

How we developed our project:

Originally, we had just planned to use data from DSCOVR as training data and data from WIND as the labels for this. After looking at the orbits using 3dView, we realized that the distance between the two satellites changed. To get around this, we used a python script to filter the data for dates when the two satellites are closest together.

Since the distance is slowly increasing, we had to base our threshold for ‘good’ data on a linear regression of a graph of distance vs. time. This threshold can be fine-tuned for the best results in the future. This python script and orbit analysis from 3dView identified 31 areas where the two satellites were the closest during the last 7 and a half years. Based on this analysis, we can find the timeframes where the data is the most reliable and use this to predict the next Carrington Event.

Another way of sorting the data is by looking for times when the two satellites are close to equidistant from the sun. We didn’t have enough time to test if this method of sorting data yields better results than our first method. This way of filtering the data can be implemented in the same way as our first one.

We also had to come up with a solution for aligning the data from DSCOVR and WIND. WIND only reports data around every 92 seconds, and DSCOVR reports data every second. The time intervals between data reports from WIND are also not consistent. To get around this, we had another python script that would check data from both satellites and merge them so that they correspond correctly.

We used data from WIND and DSCOVR that were provided by NASA. DSCOVR provided the training data, and WIND provided the labels for the training data. We used 3DView CDPP to do analysis on the orbit of the two satellites, which we used for filtering our training data.

We also used Goddard Space Flight Center - Space Physics Data Facility - CDAWeb Data Explorer, NOAA, and IRAP (Institut de Recherche en Astrophysique et Planétologie) to gather data for our project. To access the data, we started with the March 10, 2015 DSCOVR and WIND data set provided by NASA. We also researched how the instruments on DSCOVR worked, such as the Faraday Cup on NOAA's DSCOVR page.

However, we still encountered complications in visualizing the data, so we also used CDAWeb to see visual plots of the data to have a broader sense of what we were dealing with.

We also did the same method for the WIND satellite data. However, we noticed that the WIND and DSCOVR time data are not in sync. This was an extra challenge that we decided to tackle before we started our actual neural network. In addition, we used IRAP's CDPP 3D Viewer to visualize the orbits of the two spacecraft.

Our initial ideas started with brainstorming about what to use a neural network pipeline to accomplish. We thought about ways to predict when solar storms would happen, and we decided to create a neural network that learns to predict when there will be a spike in the solar wind.

The journey carries on with us settling down with the plans to make a network that uses machine learning to “sense” whether or not there is a spike in solar wind activities. Even though we are experienced in coding, machine learning algorithms proved to be a significant challenge since this was the first time we attempted to make one.

On the second day, everything was going well since we had our data collected and a plan organized. However, we also discovered a different way of filtering the data which was mentioned above.

Due to the short time we had left, we didn’t have time to make new scripts to filter the data based on the distance to the sun instead of the distance between the two satellites. In the end, it was still a fruitful experience, but if we could change something, we would go back and gather data using both methods that we designed and evaluate which one would give better results.

Links (Data):

• WIND Spacecraft data - NASA
• WIND Spacecraft Info - NASA
• DSCOVR Spacecraft data - NASA
• CDA web - Goddard Space Flight Center  - NASA
• IRAP - CDPP 3D Viewer  - IRAP (Institut de Recherche en Astrophysique et Planétologie)

Wind Solar Wind Experiment (SWE) 92-sec Definitive Solar Wind Proton Data

https://hpde.io/NASA/NumericalData/Wind/SWE/Definitive/PT92S

Ogilvie, K. W., Fitzenreiter, R. J., Lazarus, A. J., Kasper, J. C., & and Stevens, M. L. (2021). Wind Solar Wind Experiment (SWE) 92-sec Definitive Solar Wind Proton Data [Data set]. NASA Space Physics Data Facility. https://doi.org/10.48322/nasd-j276. Accessed on 2022-October-2.

Software used:

• Google Colab
• Repl.it
• Python
• Blender
• Wix 
• Topcat
• CDPP 3D viewer

#software #Carrington #machinelearning #AI #neuralnetwork