- Google used Deep Neural Network to analyze Kepler’s data.
- They found two new exoplanets – Kepler-90i and Kepler-80g.
- The model they’ve developed is about 96% accurate in distinguishing planets and non-planets.
For many centuries, people have looked up at stars, noticed some patterns and recorded observations. One of the early things discovered in space was planets, which the Greeks called ‘wanderers’ or ‘planētai’ for their irregular movements. Gradually we came to know our solar system hosts several planets revolving around the Sun.
With the help of modern technologies like digital camera, space flight, telescope optics and computers, we can extend our knowledge beyond our own solar system and detect/identify planets thousands of light years away from Earth. They are called exoplanets – part of another solar system far in space.
However, finding exoplanets is an extremely difficult task. Unlike their host stars, they are small, cold and dark. At present, we use machine learning techniques to accurately spot exoplanets. One of such techniques is used by Google, and they detected two extoplanets, named Kepler 90i, orbiting around a yellow dwarf star Kepler 90, and Kepler-80g revolving around Kepler 80. Let’s find out how did they do it.
Data Used
The primary method to search exoplanet is to analyze a vast amount of data captured by NASA’s Kepler Space Telescope using both manual analysis and automated software. In 4 years, the telescope observed around 200,000 stars, capturing an image every half an hour. Kepler’s sole scientific instrument, Photometer continuously monitors the brightness of more than 145,000 main sequence stars in a fixed view field. These data are sent to Earth, and then deeply examined to identify exoplanet’s periodic dimming caused by revolving around their host star.
This all generates about 14 billion data points, which further translate to approximately 2 quadrillion possible planet orbits. Even the most powerful computer takes an extremely long time to process these massive amounts of data. In order to make this process faster and more efficient, Google used Deep Learning tools and techniques.
The Machine Learning Approach
Machine learning is the form of artificial intelligence that teaches computer to recognize specific patterns. It is specifically helpful in making sense of large volumes of data. Here, the idea is to allow machines learn by training and examples rather than programming it with particular rules.
Image credit: NASA
Deep learning, which is a type of machine learning, uses computational layers to create progressive complex features that are useful for classification problems. For instance, a deep picture classification model may first recognize simple edge features that can be further used to detect corners and curves, until the final feature layer of the model can distinguish between complex objects.
Deep neural networks (type of deep learning model) have become the state-of-the-art in several tasks, including image classification. Most of the times it performs better than the models developed with hand designed features. A neural network is trained to minimize a cost function that measures how far its predictions are from the training set’s true labels.
The Google AI team used a dataset of over 15,000 Kepler signals to create a TensorFlow model for distinguishing planets from other celestial body. To do this, the system had to detect and recognize actual planet’s pattern vs patterns caused by other bodies like binary stars and starspots.
They developed a deep neural network for automatically examining Kepler threshold crossing events (TCEs – detected periodic signals, which might be consistent with transiting planets). The model uses light curves as inputs and is trained on a set of human-classified Kepler TCEs.
The views of input are fed via separate convolution columns – a successful method in previous image classifications. It’s capable of distinguishing space bodies with decent accuracy – the subtle differences between actual transiting exoplanet and false positives like instrument artifacts, eclipsing binaries and stellar variability.
Reference: Harvard.edu
Results
When the model was tested on signals, it correctly distinguished the signals that were generated by planets and other non-planets, with an accuracy of 96 percent. Moreover, 98.8 percent of the time it ranked plausible planet signals higher than false positive signals.
Read: NASA To Use Artificial Intelligence For Space Communication Networks
To narrow the search, they observed 670 stars that already have two or more exoplanets. While processing, they found two new exoplanets – Kepler-90i and Kepler-80g. Kepler 90i planet revolves around the Kepler-90, a star previously known to host 7 transiting planets. Whereas, Kepler-80g is the part of 5 planet chain around Kepler-80 star, with an orbital period nearly matching the prediction by 3-body Laplace relations.
Nearly 13 percent bigger than Earth, Kepler-80g (the outermost planet in its system) has an orbital period of 14.6 days and 89.35 +0.47- 0.98 degree of inclination.
Image credit: Google blog
Kepler-90i is 34% larger than Earth with an orbiting period of 14.45 days. It’s 2,545 light years away from Earth in the constellation Draco. It is located between Kepler-90c (8.7 days) and Kepler-90d (59.7 days) with an extremely hot surface temperature – 436 °C.
What’s Next?
When it comes to the possibilities of the deep neural network, sky is the limit. Out of 200,000 stars, the model is used to search only 670 of them. There may be hundreds of thousands of exoplanets still undiscovered in Kepler’s data. The new techniques like deep learning will help astronomers and physicists uncover things that are beyond human reach.
Read: Google’s Artificial Intelligence Creates an AI That Beats Human Code
This model could be modified in the future to improve its accuracy and decrease known types of false positives. For example, we can
- Increase the training set incorporating simulated data or unlabeled data (the current model uses only about 15,000 labeled examples).
- Improve fattening routine to reduce number of signals due to stellar variability that are classified as likely planets.
- Add some form of centroid information into input representation to improve system’s ability to classify transits that occur on a background star rather than the target star.
- Split the local view into several segments to allow system to analyze the consistency of transits between different segments of the dataset.
