Challenging Limitations: Using Deep Learning, Time Series Analysis, and Statistical Methods for Noise Reduction to Develop an Innovative Approach to Exoplanet Candidate Detection Using Earth-Based Telescopes
ISEF · 2018 Physics and Astronomy Fourth Award
Overview
The current limitations of our efforts toward exoplanet discovery can be attributed largely to the immense cost of the observational tools needed to look for signs of exoplanetary transits of target stars. Seeking to increase the accessibility of astronomy in a way that would heighten the rate at which new exoplanet candidates are identified, I attempted to engineer a comprehensive system for the observation and identification of possible new exoplanet candidates in areas not currently being explored by space telescopes. I proposed that this can be done by making observations using an existing worldwide network of earth-based telescopes rather than the few space telescopes currently available to us. To address the issues of atmospheric noise and periodic gaps in observation windows that arise when working with data from earth-based telescopes, a three-tiered approach was adopted to develop the analytical software. First, a convolutional neural net (CNN) was created to differentiate between transit and non-transit light curves generated using a Gaussian distribution of stellar parameters obtained from analysis of Kepler light curves translated into SIFT image descriptors. Then, 12-hour segments were deleted from the data to simulate daytime periods in which an Earth-based telescope cannot make observations. The gaps were then filled in using Gaussian Local Level and ARIMA models. Noise was artificially added to the Kepler observations to simulate observing conditions on earth, and software was developed to remove that noise through a process involving Gaussian Local Level Models and Durbin-Koopman simulations. The result of this process was software which classified transit light curves transformed under simulated earth-based observation conditions with 88% accuracy.
Awards (2)
- Fourth Award of $500 $500
- NASA: Second Award of $750 $750
Competition history
- ISEF 2018
Resources
Related projects
ISEF · 2025
Innovative Machine Learning Approaches for Exoplanet Detection and Anomaly Identification: Applying Computational Techniques to Uncover Hidden Patterns in TESS Data
ISEF · 2025
Optimizing Exoplanet Detection: A Comparison of Neural Networks and Traditional Machine Learning Models on Light Curve Data
ISEF · 2023
A Novel Method in Improving the Accuracy of Exoplanet Classification Using Machine Learning
ISEF · 2021
A Mathematically Driven Physical Analysis into Exoplanet Detection Confirmation Surveys Using Bayesian Inferential Statistics, Machine Learning/Linear Algebra Sklearn and TensorFlow Techniques, the BATMAN Python Programming Library, 3D Printing Techniques, and a Custom-made Python Processing Pipeline for Image Analysis and Lightcurve Detection
Closest projects by meaning, across every fair and year in the corpus.
Source: Regeneron International Science and Engineering Fair