1. AI & Space
FRANCK MARCHIS
Senior Planetary Astronomer at the SETI Institute
Chair of the Exoplanet Group
Chief Scientific Officer at Unistellar
With contributions by the FDL Team
2. Overview of today’s discussion
1. SETI Institute & Space Science
2. NASA & FDL
3. A future with AI in Space
Outline
3.
4.
5.
6.
7. SETI Institute and Astronomy
3 ways to answer to this question
Technological Signal
• Radio SETI
• OSETI
Intelligent
Nearby & distant
Life in Our Solar System
• Ocean of Europa
• Underneath Mars Surface
Not-intelligent
Nearby
Life on Exoplanets
• Remote sensing observations
• Biological signatures
• Technological artifacts
Intelligent and Not-Intelligent
Distant
Are We Alone
9. It's not often that NASA asks you to come to Silicon
Valley to save the world using AI…
But that’s exactly what happened in summer 2016 to 12 researchers
from around the world. and again in 2017… and soon 2018
10. ABOUT FDL
GOAL: CLOSE KNOWLEDGE
GAPS BY MATCHING DOCTORAL-
LEVEL TALENT FROM THE
PLANETARY SCIENCES WITH
PEERS FROM
THE MACHINE LEARNING
COMMUNITY.
11. “To find all asteroid threats to
Human Populations and know
what to do about them”
15. 3)’What is the best choice of
technology to make a successful
deflection of an Asteroid that
poses a threat to Earth?’
16. ‘What are they made of?’ ‘What shape are they?’
‘What is the best way
to deflect them?’
(Better Data Gathering) (Better Insights) (Better Decisions)
17. “Only 31 meteorites have
ever been found that can be
linked to a source orbit”
AI for Planetary Defense: ‘What is it made of?’
NASA needs more fragment recovery.
Can AI help?
18. AI For Planetary Defense
Fragment Recovery
“Only 31 meteorites have ever been found that can be linked to a source orbit”
25,000 training images of meteorites
six deep learning models (3 GoogleNet, 2
AlexNet + 1 combination) along with a 15
million image library.
•Result: an automatic meteorite detection
system, all driven by a user-friendly app for
use in the field.
Validation accuracy > 99.9%
21. •Asteroid shapes are critical for asteroid
deflection techniques - as any mitigation plan
needs to know the center of mass.
•Should an object be too close to shift, shape,
density, internal structure are critical for
understanding the potential for damage and
planning effective disaster response.
AI for Planetary Defense:‘What shape is it?
22. FOUR
WEEKS
This is currently a laborious
manual process that takes
a trained practitioner
around four weeks.
AI for Planetary Defense:‘What shape is it?
23. • Create 546,000 synthetic radar
images from DAMIT and JPL shape
database (1,620 models)
• This approach may enable a rapid understanding of the shape of an asteroid - while it is still
being tracked by radar, even as an incoming object.
• Automatically determine the spin
axis angles by automated calls of the
existing SHAPE software using
Bayesian Optimization (4-6h
computer time)
• Used a 3D-VAE (Variational Auto-
encoder) to generate plausible 3D
voxel shapes of asteroids
AI for Planetary Defense: ‘What shape is it?
24. 3D SHAPE MODELING
THE VARIATIONAL AUTO-ENCODER (VAE)
GANs attract global attention
With Deep Mind’s ‘Alpha Go”
In 2017!
25. 3)’What is the best choice of
technology to make a successful
deflection of an Asteroid that
poses a threat to Earth?’
26. AI for Planetary Defense: Deflecting Asteroids?
‘What is the best choice of technology to make a successful
deflection of Asteroid that poses a threat to Earth?’
Nuclear
Device
?
27. Why a Deflector Selector? A TRAINING SET BASED ON 1.5
MILLION ORBITAL SIMULATIONS OF
THREE DIFFERENT KINDS OF
DEFLECTION TECHNOLOGY.
“There isn’t a tool of this
sophistication available to the
Planetary Defense community.”
Astronomer, JL Galache from the IAU’s Minor Planet Center (& FDL Mentor)
AI for Planetary Defense: Deflecting Asteroids?
28. 1.5 million orbital simulations were used train a
decision tree to select a set of effective technologies
(Nuclear Device, Gravity Tractor or Kinetic
Impactor) for a given hazardous object.
Once trained, it had an accuracy of 98% for
determining which technology would produce a
successful deflection.
A Machine Learning
Decision Tree…
AI for Planetary Defense: Deflecting Asteroids?
29. •The most effective technology predicted by the
decision tree is the nuclear explosive, due to the
high ΔVs it can impart and its instantaneous
effect.
•This important work will help inform strategic
decisions which deflection technologies should
be prioritized, and what asteroid characteristics
are the most important to be known in advance
of taking action.
Greenberg et al. Acta Astronautica 2018
Deflector Selector: Result
30. PLANETARY
DEFENSE
2016
3 PROBLEM AREAS
12 RESEARCHERS
4 PLANETARY MENTORS
2 AI MENTORS
RADAR SHAPE
MODELING
2017
HELIOPHYSICS
SPACE RESOURCES
5 PROBLEM AREAS
24 RESEARCHERS
6 PLANETARY MENTORS
6 AI MENTORS
LONG PERIOD COMETS
AI
APPLIED
Expanding FDL: The 2017 version
31. FDL 2017: RESULTS
5 CHALLENGES + WHITE PAPER
SPACE RESOURCES
HELIOPHYSICS
PLANETARY
DEFENSE
WHITE PAPER
33. PLANETARY
DEFENSE
2016
3 PROBLEM AREAS
12 RESEARCHERS
4 PLANETARY MENTORS
2 AI MENTORS
RADAR SHAPE
MODELING
2017
HELIOPHYSICS
SPACE RESOURCES
5 PROBLEM AREAS
24 RESEARCHERS
6 PLANETARY MENTORS
6 AI MENTORS
2018
LUNAR ROUTE
PLANNINGS
SOLAR STORMS /
RADIATION WARNING
LONG PERIOD COMETS
7 PROBLEM AREAS
28 RESEARCHERS
7 PLANETARY MENTORS
7 AI MENTORS
AI
APPLIED
FDL 2018
ASTROBIOLOGY
EXOPLANETS
34. FDL 2018 or FDL 3.0: Bigger, Bolder and Broader
35. Enhanced Vision for Exploration, Learning and Citizen Science.
Unistellar : making astronomy fun and popular
Enhanced Vision
technology
Beautiful images
Light pollution compensation
Portable and
autonomous
Fits in an urban
backpack
Connected
Science campaign
Data Sharing
Automatic Field
Detection
Educative and Interactive
37. Future of Ground Based Astronomy
Extremely Large telescopesAll the Southern Sky, every week: LSST
20 TB of data per night
20 million of asteroid
100 alerts per day
AI to process
AI to alert
AI in optical and mechanical design
AI in data analysis
38. Future of Space-Based Astronomy
The picture of another Pale Blue Dot: Project Blue
Searching for exoplanets: TESS
27 Gb of data per night
200,000 stars
AI to process
AI to Alert
2 years of continuous observations of Alpha Centauri stars
AI to process the data
We want to make sure there’s plenty of time for discussion, but of course if you have a question or reflection feel free to jump on in.
2 trillion of galaxies
100 billion stars
Unambiguous earliest evidence for life on Earth dated to 3 billion years
But that’s exactly what happened last Summer to 12 researchers from around the world.
FDL is an Applied Research Accelerator: that is, the application of emerging technology to known ‘know-how’ gaps; by pairing the talent from academia and the private sector.
When we talk about planetary defense, we mean Asteroids; or ‘Near Earth Objects’ and this project was a response to the White House Challenge to find all asteroid threats to Human Populations…AND KNOW WHAT TO DO ABOUT THEM.
So the thing they never mention in the movies - you need to know three things in advance.
What is it made of?
What shape is it?
What is the best choice of technology to make a successful defection?
Problem choice is key too. We’ve endeavored to keep the problems related to a certain extent, so that the approaches and ideas created by one team, inspire the others. This slide shows the challenges from 2016, which also show the spread of solutions where AI can be applied: “Better Data Gathering” / “Better insights” / and “Better decisions”
This is a very rare object - discovered this year; it’s a meteorite where the fireball - the shooting star - was tracked by fireball cameras. This has only happened 30 times. (Where a rock and the incoming trajectory are linked)
The next thing we need to know is, what shape is it?
When NEOs come near the Earth, it is possible to image them using radar. From these sparse 2D images, the 3D shape can be discerned…
Why is this important? Any defection plan needs to know the object’s shape and center of mass; should an object be too close to move, knowing it’s shape can help with disaster response.
but this is currently a laborious manual process.
It may be possible to discern the shape of an object, even as an incoming object.
Before we could do anything else, we needed data.
Of course, in an ideal world, we would have lots of data of many different objects which humanity had tried deflecting.
Actually maybe that’s a not-so-ideal world.
Anyway, since we don’t have a sample of these kinds of objects, we had to resort to creating our own data with simulations.
This year the team improved this approach by adding a GAN to the process, modifying the VAE to get further levels of precision.
Lastly, now we know what it’s made of and its shape, what is the best technology to make a successful deflection?
Although have been making strides in engineering solutions to NEO deflection, the best technology assessment was - up until FDL - based on only a handful of orbits.
The team radically improved the resolution, by running 1.5 Million orbital simulations where NEOs where on a collision course with Earth and successfully deflected - to create training data for a decision making tool…
A machine learning decision tree… once trained, it had an accuracy of 98% in determining which technology would produce a successful deflection…
The most effective tool we have, then - as Bruce Willis demonstrated - is the nuclear device. However I know he didn’t do the math. Now we have the math.
We’ve extended this concept of an ‘envelope of expertise’ to extend and branch the work of the preceding year.
In 2017 we revisited PD with the same mentor team, adding Helio and Space Resources.
We’re looking to build on the work of 2017 and branch out again - there’s appetite to translate the ‘foggy’ work done in Helio this year into quests. And of course we’d like to continue pushing the possibilities in Planetary Defense.
Apollo Guidance Computer (AGC): It had approximately 64Kbyte of memory and operated at 0.043MHz. The scientist in charge of the software development program for the Apollo Guidance Computer was Margaret Hamilton, Director of the Software Engineering Division of the MIT Instrumentation Laboratory. Curiously, the world's first computer programmer was also female