NORMA eResearch @NCI Library

Discovery 2: An analysis of Machine Learning methods to predict exoplanet candidates: Technical Report

Flanagan, Jonathan (2022) Discovery 2: An analysis of Machine Learning methods to predict exoplanet candidates: Technical Report. Undergraduate thesis, Dublin, National College of Ireland.

[thumbnail of Bachelor of Science]
PDF (Bachelor of Science)
Download (3MB) | Preview


This project examines the performance of Convolutional Neural Networks and Capsule Networks in the problem domain of classifying exoplanet candidates using light fluctuation readings from NASA telescope data. NASA telescopes record light intensity readings from observed stars, if a planet is orbiting one of these stars on the same visual plane as the telescope, an identifiable dip in light intensity is created. This type of reading is known as a transit event and can take different forms depending on the number and size of planets orbiting a particular star. This method of detection makes up 76.82% of the 5,030 exoplanets discovered.

Convolutional Neural Networks were identified as the state of the art machine learning model used for this type of classification problem, and although Capsule Networks have primarily been designed for computer vision tasks and are a relatively new concept, the project aimed to test their ability in this problem domain.

Data retrieved on confirmed exoplanets as well as false positive candidates from the Mikulski Archive for Space Telescopes was used in the analysis and the results showed the Capsule Network performed better in terms of accuracy with 0.96 compared to the Convolutional Neural Network at 0.95. The Capsule Network also performed better in precision with a 1.0 score while the Convolutional Neural Network returned 0.91. The Convolutional Neural Network scored better in recall though with 0.90 versus the Capsule Network at 0.85.

Comparing AUC scores, the Convolutional Neural Network scored best with an AUC of 0.937 while the Capsule Network scored 0.928. Conversely, comparing F1 scores of the models showed that Capsule Network scored best with 0.92 versus the Convolutional Neural Networks score of 0.90. A McNemar’s statistical test returned a p-value of 0.185 concluding that there was no statistical difference between the proportion of errors between both models.

Item Type: Thesis (Undergraduate)
Subjects: Q Science > QA Mathematics > Electronic computers. Computer science
T Technology > T Technology (General) > Information Technology > Electronic computers. Computer science
Q Science > QA Mathematics > Computer software
T Technology > T Technology (General) > Information Technology > Computer software
Q Science > QB Astronomy
Divisions: School of Computing > Bachelor of Science (Honours) in Computing
Depositing User: Clara Chan
Date Deposited: 30 Aug 2022 12:05
Last Modified: 30 Aug 2022 12:05

Actions (login required)

View Item View Item