Chasing Exoplanets using AI
Used the data from the Kepler telescope to train a convolutional neural network that uses the transit method to detect the presence of exoplanets in space. This project is a replication of the original work of Google Brain team Engineer Chris Shallue.
Introduction
People have been looking up at the sky for thousands of years, recording observations and noticing patterns. Planets, which the Greeks called "plantai," or "wanderers," for their seemingly random movement through the night sky, were among the first objects observed by early astronomers. People have known for centuries that the Earth and other planets in our solar system orbit the sun, which is a star like many others.
Today, we can extend our understanding beyond our own solar and find planets surrounding other stars using technology such as telescope optics, space flight, digital cameras, and computers. The study of these planets, known as exoplanets, allows us to delve deeper into some of humanity's most profound questions about the cosmos.
Is there anything else out there? Are there any other planets or solar systems that are similar to ours?
Though technology has made the search easier, locating exoplanets is still a difficult task. Exoplanets are cold, tiny, and dark in comparison to their host stars, making them as difficult to notice as a firefly flying next to a searchlight from thousands of miles away. However, we've recently made some progress thanks to machine learning.
Using both automatic tools and manual analysis, astrophysicists seek for exoplanets by examining vast volumes of data from NASA's Kepler spacecraft. For four years, Kepler photographed around 200,000 stars every 30 minutes, resulting in 14 billion data points. There are almost 2 quadrillion potential planet orbits based on those 14 billion data points!
It's a massive quantity of data for even the most powerful computers to comprehend, making it a difficult and time-consuming operation. We used machine learning to help this process go faster and more efficiently.
Machine learning is a technique for teaching computers to spot patterns, and it's very helpful when dealing with massive amounts of data. Instead of programming a computer with precise rules, the main concept is to let it learn by example.
We built a TensorFlow model to differentiate planets from non-planets using a dataset of over 15,000 tagged Kepler signals. It has to distinguish between patterns created by planets and patterns caused by other objects such as starspots and double stars in order to do so. Our model successfully identified which signals were planets and which signals were not planets 96% of the time when we tested it on signals it had never seen before. As a result, we knew it worked!.
Poster Design Competition
123-456-7890
Methodology
Let's start with how the Kepler telescope's data is used to determine the presence of a planet. A light curve is a graph that depicts the brightness of a star (as measured by Kepler's photometer) through time. When a planet passes in front of a star, it temporarily blocks some of the light, causing the measured brightness to fall and then rise again, resulting in a "U-shaped" dip in the light curve.
Other astronomical and instrumental events, such as binary star systems, starspots, cosmic ray strikes on Kepler's photometer, and instrumental noise, might cause a star's estimated brightness to decline.
Scientists employ automatic algorithms (such as the Kepler data processing pipeline) to find signals that could be created by planets in Kepler data, and then manually follow up to determine whether each signal is a planet or a false positive. To prevent being overwhelmed by too many signals, the scientists apply a cutoff to the automated detections: those with signal-to-noise ratios above a certain threshold are considered worthy of further investigation, while all detections below the threshold are rejected.
There's an enticing incentive: it's likely that some potentially habitable planets like Earth, which are small and orbit around dim stars, are hiding just below the usual detection threshold – there could be hidden jewels in the Kepler data that have yet to be uncovered!
