This document discusses space weather and solar-terrestrial interactions. It introduces B-STING, an open-source tool that uses machine learning to predict the Kp index, which measures geomagnetic activity levels. It evaluates several machine learning models, finding that gradient boosting achieves the best prediction of Kp values up to 3 hours ahead. Feature analysis reveals current Kp index and solar wind parameters are most important for predictions. B-STING utilizes scientific data and machine learning algorithms to better understand and forecast space weather impacts.

1. SPACE WEATHER MISSION 02
SOLAR-TERRESTRIAL
INTERACTIONS

2. SPACE WEATHER: SOLAR TERRESTRIAL INTERACTIONS
MEET THE TEAM

5. Aurora T
SPACE WEATHER: SOLAR TERRESTRIAL INTERACTIONS
SPACE WEATHER
Aurora
GPS
Satellites
Astronauts
Power Grid
Aviation
Credit: NOAA/SWPC

6. SPACE WEATHER: SOLAR TERRESTRIAL INTERACTIONS
SATELLITES
Credit: NASA/ESA

7. SPACE WEATHER: SOLAR TERRESTRIAL INTERACTIONS
DATA SOURCES
GEOMAG DATA SOLAR WIND DATA

8. SPACE WEATHER: SOLAR TERRESTRIAL INTERACTIONS
PROBLEM DEFINITION
B-STING: Solar Terrestrial Interactions Neural Network Generator
A data-driven, open source tool for space weather forecasting.
Predicts a commonly used index, called Kp, that captures geomagnetic disturbances.
Value Proposition:
Open Access Science Data x AI and ML frameworks
=
Better Scientific Insights + Better Predictions
SPACE WEATHER: SOLAR TERRESTRIAL INTERACTIONS
PROBLEM DEFINITION

9. Planetary Kp Index
(Bartels, 1938)
Kp Index - refers to a range
of geomagnetic activity
levels within a 3-hr interval
each day (in UT)
Kp varies from 0 to 9;
quasi-logarithmically
SPACE WEATHER: SOLAR TERRESTRIAL INTERACTIONS
Kp INDEX

10. SPACE WEATHER: SOLAR TERRESTRIAL INTERACTIONS
Kp INDEX

11. G-Scale Kp Activity Level Occurrence Frequency
G0 4 and lower Below Storm
G1 5 Minor Storm 1700 per cycle (900 days per cycle)
G2 6 Moderate Storm 600 per cycle (360 days per cycle)
G3 7 Strong Storm 200 per cycle (130 days per cycle)
G4 8 Severe Storm 100 per cycle (60 days per cycle)
G5 9 Extreme Storm 4 per cycle (4 days per cycle)
SPACE WEATHER: SOLAR TERRESTRIAL INTERACTIONS
STORM LEVEL & Kp INDEX

12. Date Event Level
1 Sept 1859 Carrington Event
widespread disruption of telegraph
Extreme
13 March 1989 Hydro-Quebec
9 hour black out
Severe
20/21 Jan 1994 Anik-E1 and Anik-E2 failed
Disrupted TV and computer transmission
Moderate
14 July 2000 Bastille Day Event Extreme
31 October 2003 Halloween Events
Affected airlines, caused power outages,
damaged transformers, led astronauts on
ISS to take shelter
Extreme
The Sun affects the
near-Earth environment,
which leads to disruptions
in our technological
systems.
SPACE WEATHER: SOLAR TERRESTRIAL INTERACTIONS
SPACE WEATHER EVENTS

13. SPACE WEATHER: SOLAR TERRESTRIAL INTERACTIONS
MACHINE LEARNING
Recurrent Neural Nets (RNNs)
(Geomag + Solar Wind)
Ensemble
(Geomag + Solar Wind + Kp)
Long Short Term Memory
(LSTM)
Gradient Boosting
AdaBoosting
Random Forest
Gaussian Process
Bagging
Extra Trees
Project 1 Project 2

14. Special flavor of Recurrent Neural Networks (RNNs), capable of learning long term
dependencies.
RNNs have loops Chain-like nature of RNNs make them suitable for
time series data
input
memory
block
output
SPACE WEATHER: SOLAR TERRESTRIAL INTERACTIONS
LONG SHORT TERM MEMORY RECURRENT NEURAL NETWORK

15. SPACE WEATHER: SOLAR TERRESTRIAL INTERACTIONS
LONG SHORT TERM MEMORY RECURRENT NEURAL NETWORK
Attempted to predict Bz on Earth using Bz@L1 and solar wind speed.
Managed to predict Bz@L1 and solar wind speed using Bz on Earth.

16. Each tree is an estimate of the solution to the overall problem.
Trees are reweighted after each iteration of training.
Allows each weak tree to contribute to an overall strong solution.
SPACE WEATHER: SOLAR TERRESTRIAL INTERACTIONS
GRADIENT BOOSTING
Find an ensemble of trees so that, when added together, minimizes a loss function

17. SPACE WEATHER: SOLAR TERRESTRIAL INTERACTIONS
GRADIENT BOOSTING

18. SPACE WEATHER: SOLAR TERRESTRIAL INTERACTIONS
GRADIENT BOOSTING RESULTS
Prediction of Kp 3 hours ahead

19. SPACE WEATHER: SOLAR TERRESTRIAL INTERACTIONS
METRIC: MEAN SQUARED ERROR
ML method 1h ahead 3h ahead 6h ahead
Persist 0.007 0.020 0.025
Mean 0.046 0.046 0.046
Median 0.048 0.048 0.048
Gradient Boosting 0.007 0.015 0.021
Adaptive Boost 0.012 0.018 0.032
Extra Trees 0.009 0.021 0.027
Random Forest 0.015 0.015 0.026
> 95%
confidence level

20. SPACE WEATHER: SOLAR TERRESTRIAL INTERACTIONS
FEATURE DISCOVERY
Most important: Current Kp index
Other important predictors:
- Solar wind magnetic field strength and Bz,
- Solar wind speed and proton density,
- Unexpected Result: N-S component of the geomagnetic field at
low latitude stations (Guam, Hawaii, Puerto Rico). This points to
the importance of the magnetospheric ring current.
This plot shows the relative importance of the physical parameters for Kp prediction.
Machine learning extract important physical parameters without a
priori knowledge of the system.

21. Utilized big data (well calibrated, science quality)
Used state-of-the-art AI and ML algorithms
Developed a tool that can predict geomagnetic index Kp
Cross-validated physical assumptions used in scientific research
ML extracted important physical parameters without a priori knowledge of the system!
B-STING!
https://github.com/nasa-fdl/solar-terrestrial
An open source python module for Heliophysics data analytics,
built on the latest ML/AI frameworks.
This is just the beginning...