High-Level Project Summary
Our challenge was to design an algorithm to improve solar wind data quality from the satellite DSCOVR in order to make better predictions of solar activity events like solar flares, coronal mass ejections and solar storms. We used a variety of techniques and ultimately designed a neural network to predict wind speed, density and temperature from the magnetic field data provided. More accurate predictions and data collection allows for better mitigation of potential solar hazards. In fact, we developed a series of infographics with the different categories of solar events with information on what to expect and a small aurora borealis predictor based on solar wind data.
Link to Final Project
Link to Project "Demo"
Detailed Project Description
In this project, we have developed a new, powerful tool to compute parameter response functions. We have demonstrated the proposed Neural Network to derive the speed, temperature, and density of particles. The NN was trained on a specific range of solar wind parameters and it works on a dataset of two satellites, Wind and DSCOVR. Our technique is intended to provide more accurate input parameter calculations than previous fitting techniques.
Another Carrington event in today’s world would lead to electric systems collapsing and would cause damages which could take years or even decades to repair. It is crucial to develop and maintain systems that can provide early warnings for these kinds of events. Their information would eventually be priceless.
In our attempt to improve the data quality from DSCOVR, we have designed a Sequential Neural Network to process magnetic field vector data obtained from the Wind and DSCOVR satellites that ultimately predicts wind ion parameters (velocity, temperature and density). The network would have the following properties:
- Multiple Vector Input
- 3 layers
- 3 neuron output (one for each wind parameter)
The ‘chi-sq’ loss function seems to be sufficient for this project, but more sophisticated functions could be used. We determined that using 6 to 9 months of data to train the network is ideal, since it considers a wide range of solar activity and is not unnecessarily long. Since the time and computer power required to train the proposed network exceeds the time-frame of the hackathon, we have developed a small-scale test so as to prove our concept.
To implement the AI, we used Python and code libraries like tensorflow, keras, sklearn, pandas, amongst others. In our scaled-down model, the input is only one magnetic field vector and we trained it with data from January 2022 to February 2022. In order to reduce the amount of training parameters, we only considered one vector every 5 minutes. The resolution of the data is therefore greatly reduced. However, we managed to achieve promising results that show the potential of the tool.
As an extra, aiming to share the project with those who don’t want to dive into the details of data analysis, we did some research on the different categories of solar storms and flares, as well as the impacts they would have and measures to mitigate side effects. The output of our neural network could be fed into a program that will send alerts, warnings and instructions to the agency and/or users about peeks in solar activity.
Solar wind can also be fun! When solar activity is high, some beautiful aurora borealis can be seen in the poles. One of our ideas was to make a tool to visualize the probability of seeing auroras as well as the latitudes between which the phenomenon occurs, again, using the output of our neural network.
Space Agency Data
- The Wind Missions Magnetic Field Data Sets, BW(t): There we have all the satellite Wind data, about the intensity of the solar wind.
-https://cdaweb.gsfc.nasa.gov/pub/data/wind/mfi/mfi_h2/2022/ [NASA]
- The DSCOVR magnetic field data sets: Here we have all the satellite DSCOVR data, about the intensity of the solar wind. -https://cdaweb.gsfc.nasa.gov/pub/data/dscovr/h0/mag/2022/ [NASA]
- CDF library: We use it as a tool to be able to read the data from the satellites.
-https://www.nesdis.noaa.gov/current-satellite-missions/currently-flying/dscovr-deep-space-climate-observatory - [NASA]
- DSCOVR: Deep Space Climate Observatory: To know how DSCVRD satellite work
-https://www.nesdis.noaa.gov/current-satellite-missions/currently-flying/dscovr-deep-space-climate-observatory -- - [NASA]
- NOAA Space weather scales:To see the different density and possible consequences, of the geomagnetic storms,radio blackouts and solar radiation storms.
-https://www.swpc.noaa.gov/noaa-scales-explanation [NASA]
Hackathon Journey
With our group, we became keenly interested in saving Earth from another Carrington event. We chose this challenge since it was exciting and gave us a chance to contribute our grain of sand in order to prevent catastrophe from a potential solar storm.
We also wanted to study this type of catastrophic event and acquire the knowledge to mitigate its effects when a Carrington-like event happens again. To achieve our goal, we designed a neural network trained with data from two satellites in order to predict wind ion parameters like speed, temperature and density. This would then be used to detect if there is potentially damaging solar activity, with the aim of giving early warnings to entities and users.
References
Tags
#sun, #AI, #Carrington, #solarstorm, #neuralnetwork