logoOn 20-21 November 2018, the Toronto Machine Learning Summit (TMLS) is being held at The Carlu, Toronto. TMLS is a collaborative event for the Canadian Machine Learning community comprised of over 6,000 Machine Learning researchers, professionals, and entrepreneurs across several Bagdisciplines.

Artash (12 years) is interested in space, robotics, music, and artificial intelligence. He has been working on using machine learning to predict the risk of Asteroids colliding with the Earth. He submitted a paper to the TMLS on Space-Rex: An A.I Algorithm to Predict Risk of Asteroids Colliding with Earth. He was invited to present a poster on his project at the Summit on 20 November 2018.

explaining
Artash explaining his project – SpaceRex which uses machine learning to predict the risk of Asteroids colliding with Earth

Background to Space-Rex Project

Asteroids are rocky remnants left over from the early formation of our solar system about 4.6 billion years ago. They orbit our Sun but are too small to be called planets. Hundreds of thousands of these minor planets are gathered in the main asteroid belt between the orbits of Mars and Jupiter. Asteroids that pass close to Earth—and merit close watch—are called Near-Earth Objects, or NEOs. The current total of NEOs stands at 19191 (As of 20 November 2018) and is increasing.

Slide1

Predicting risk index of asteroids collision is a challenging exercise. Many of the Asteroids are too small to be detected from Earth in advance and it is difficult to predict their precise orbit. They can only be observed on final approach. For instance, on 15 April 2018, Asteroid 2018 GE3 swept by at half the moon’s distance Sunday, just hours after being detected. Its size was 3 to 6 times that of the space rock that penetrated the skies over Chelyabinsk, Russia, in 2013.

Current mechanisms for detecting asteroids on final approach rely on wide-field ground-based telescopes. But these telescopes are not powerful enough to detect the smaller asteroids (less than 140 meters).

Space-Rex intends to create an index to predict the risk of an asteroid colliding with Earth using data from NASAs’ Centre for Near Earth Objects (CNEOS).

Slide8
Data used by the Space-Rex project

Enter Machine Learning

Artificial Intelligence can be useful in this case as the amount of data coming from Sky Surveys and ground-based telescopes are rising manifold. But the number of astronomers to look for patterns are not. However an A.I can be trained to look for patterns in massive data in real time.

Slide3Space-Rex uses a Feed Forward Neural Network to solve this challenge. In Feedforward Neural Networks the information only travels forward in the Network, first through the input nodes, then through the hidden nodes, and finally, through the output nodes –  the connections between these units do not form a cycle. They are primarily used for supervised learning in cases where the data to be learned is neither sequential nor time-dependent.

Slide4
Input Layer for Space-Rex

The input layer to this Neural Network constituted of 3 parameters of the Asteroid:

  1. The Velocity of the Asteroid: An asteroid moves fast – with an average speed of 25 kilometers per second and there is uncertainty in their orbital path. Multiple observations are needed to trace their path accurately.
  2. The Diameter of the Asteroid: Asteroids range in size from Ceres—the largest at about 945 kilometers in diameter – to bodies that are less than 10 meters across. The total mass of all the asteroids combined is less than that of Earth’s Moon. The bigger the Asteroid, the greater the devastation it can cause.
  3. Apparent Magnitude (Visibility): Most asteroids are irregularly shaped and have a different albedo (reflectiveness). The asteroids also rotate, sometimes quite erratically, tumbling as they go, making it difficult to predict their orbit. The Asteroids which have low apparent magnitude are difficult to detect. About two-thirds of all asteroids are thought to be C type asteroids. These carbon-rich asteroids are very dark, with an average albedo of about 0.06. The S (silicates) type asteroids are considerably brighter with an average albedo of 0.16 while M (metal) type asteroids have an average albedo of 0.19.
Slide5
The Feed Forward Neural Network used in the Project

The output layer of the Neural Network was the Risk Prediction Index modeled on NASA’s Palermo Technical Impact Hazard Scale. It compares the likelihood of the detected potential impact with the average risk posed by objects of the same size or larger over the years until the date of the potential impact. This average risk from random impacts is known as the background risk. The Index is logarithmic and can have positive and negative values. Positive values indicate higher than the background risk for a collision while negative values indicate lower than the background risk for collision.

Slide6
Output Layer for Space-Rex

Conclusions from Space-Rex

Interestingly none of the Asteroids included in the deployment database for the A.I Algorithm came up with a positive risk index. It indicates that the risk of an asteroid colliding with Earth is very small but non-zero. The index followed a normal distribution centered at-risk index of -4. It implies that the current risk is ten thousand times less than the average risk from random impacts.

Slide9
Asteroid Risk Index follows a Normal Distribution. 90% of Asteroids have risk index less than -3 (1000 times unlikely to occur than a random background event)

As new asteroids are discovered they fit nicely into this normal distribution which bodes well for the Earth’s future. It means that it is very likely that the risk of collision posed by yet to be discovered Asteroids is very low (but non-zero).

More observations need to be made especially of smaller sized Asteroids to improve the accuracy of the model. This Model assumes significance as we expand on Space Exploration and plan for Lunar Gateway. It can also be expanded to include space junk created by older and non-functional satellites.

Slide10
Conclusions from the Space-Rex project

A Palermo Scale value of -2 indicates that the detected potential impact event is only 1% as likely as a random background event occurring in the intervening years, a value of zero indicates that the single event is just as threatening as the background hazard, and a value of +2 indicates an event that is 100 times more likely than a background impact by an object at least as large before the date of the potential impact in question. (Source: NASA Centre for Near Earth Objects (CNEOS) https://cneos.jpl.nasa.gov/sentry/palermo_scale.html)

nick frosst
Explaining his project to Nick Frosst, Researcher at Google Brain
Patrick Hall.JPG
With Professor Patrick Hall, York University, Toronto – could we use a similar technique for quasar detections?

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s