OpenPOWER NASA IBM AI Object Automation

1. AI Research and
OpenPOWER at the NASA
Frontier Development Lab

2. Some context...

3. IBM and the SETI Institute
l 4.5TB data coming
from the beamformers
every hour
l Data searched for
narrow-band signals
using FFT in custom
hardware
l “Blind” to other types
of potential signals of
interest
l Most data is dumped –
only the data with
detected signals is
saved for later analysis

4. IBM and the SETI Institute
setiQuest/ML4SETI

5. IBM and the SETI Institute
http://bit.ly/ibm-trappist
Kepler Transit Observations – only works for the small fraction of
star systems with orbital planes that are perfectly edge on...
ü û

6. Applied artificial intelligence research accelerator that combines the
AI capabilities of NASA, academia, and private sector companies in
support of NASA’s science priorities and mission goals.

7. A few examples from 2017...

8. •The FDL team tackled the task of automating
task of creating 3D shape models of NEOs
from sparse radar data – important for threat
assessment and orbit fitting
•The process currently takes up to four
weeks of manual interventions by experts
using established software.
•The team demonstrated a DL model for
automation that allows NEOs to be 3D shape
generated in several hours.
•This result will hopefully allow researchers
to render 3D models of the current backlog
of radar imaged asteroids.

9. 3D SHAPE MODELING PIPELINE
hours 50 - 100 API calls
days
week

10. • Vast amounts of data collected by satellites and
observatories - largely untapped for research into how the
Sun interacts with Earth.
•The FDL team built a knowledge discovery module named
STING (Solar Terrestrial Interactions Neural Network
Generator) on top of industry-standard, open source
machine learning frameworks to allow researchers to
further explore these complex datasets.
•The FDL Solar-Terrestrial team trained an NN model that
showed the ability to predict the variability of Earth’s
geomagnetic fields in response to solar wind
•Interesting twist: the model correctly tuned into the imprint
of the magnetospheric ring currents – a subtle but
important precursors of geomagnetic storms, and an
example of unsupervised AI discovery.

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

12. •Current operational flare forecasting relies on human
morphological analysis of active regions and the
persistence of solar flare activity.
•The FDL team performed analyses of solar magnetic
complexity and deployed convolutional neural
networks to connect solar UV images taken by
SDO/AIA into forecasts of maximum x-ray emissions.
•The technique has the potential to improve both the
reliability and accuracy of solar flare predictions.

13. SPACE WEATHER: SOLAR STORM PREDICTION
SDO/AIA Image Channels
Goal:
Use AIA to forecast
GOES X-ray flux
1 hr Ahead

14. SPACE WEATHER: SOLAR STORM PREDICTION
FlareNet
Block 2
Filter 8
Texture
Block 3
Filter 7
Structure
FlareNet NN learned the importance of active regions…

15. SPACE WEATHER: SOLAR STORM PREDICTION
Solar Flare Forecast
Neural Net was able to
forecast an upper level
of X-ray flux activity.

16. The SETI Institute is a member
of
the OpenPOWER Foundation

18. Example: Lunar Resources for Permanent Outposts
(e.g. Ice water in perma-shadows, ready-made shelters, etc.)
Lava Tube Skylights v.s. Craters

19. Example: The Sun Tomorrow, Today
24 hour forecast of the appearance of the Sun
l Solar Dynamics Observatory
AIA instrument … 12 HD
images of the Sun every 12
seconds since 2010.
l Use petabytes of SDO AIA
images to build a neural net
model to produce a forecast of
what the sun will look like 24
hours into the future.
l The trained neural net model
would ingest a time-series
sequence of AIA images
leading up to the present
moment, and output an image
of the sun as it is predicted to
appear in 24 hours.
Today
Tomorrow
Forecast

20. Example: Revive the SDO “EVE” Experiment
EVE MEGS-A – Shut down in 2014 due to a defect… can we create a
NN that is a “Virtual MEGS-A” using data from other SDO readings?l Solar Dynamics Observatory EVE instrument … a
critical source of data on solar irradiance (used for
satellite orbital boost planning)
l 4 years of MEGS-A data collected before defect &
shut down (2010-2014)
l Can a NN model be trained to revive EVE to its full
potential by creating a “virtual” MEGS-A instrument
by training an AI model with historic AIA images
correlated to the 4 years of MEGS-A data that was
collected prior to its shut down? Stay tuned!
û

21. Want to participate in NASA FDL 2018?
June 25 – August 15 Mountain View California
www.frontierdevelopmentlab.org