Using AI to identify new worlds
Astrophysics Hub
Students from Highams Park School and Marina Ventikos
A research project conducted by Students from Highams Park Academy Trust, along with Orbyts Fellow Marina Ventikos from Mullard Space Science Laboratory, sought to demonstrate the significant utility of artificial intelligence in the ongoing search for exoplanets. Their approach involved a two-pronged strategy: an initial visual analysis of a dataset by the human research team, followed by the development and testing of a convolutional neural network on the identical dataset to ascertain its performance in identifying planetary signals. This project aimed to prove that AI could significantly assist humans in finding planets beyond our solar system!
This AI was trained using a dataset of 396 light curves, containing labelled examples of three classifications: Exoplanet, Eclipsing Binary, and Star. They chose to create and develop a convolutional neural network, passing inputs through a series of layers to learn features and make decisions, with the best-performing model consisting of three convolutional layers with max-pooling and three fully-connected layers. The testing dataset subsequently comprised 99 light curves, with an equal distribution of 33 light curves for each classification.
The results compellingly illustrate the AI's superior capability in exoplanet detection, notably outperforming the human researchers. The human team successfully identified 18 out of 33 exoplanets within the testing dataset, achieving an overall accuracy of 66%. Their performance is further detailed in a confusion matrix, comparing true labels to predicted labels for exoplanets, eclipsing binaries, and stars.
In stark contrast, the developed AI system accurately recovered 30 out of the 33 exoplanets, demonstrating a notable overall accuracy of 89%. The AI's confusion matrix and metrics also show higher precision, recall, and f1-scores across all classifications compared to human results, confirming its enhanced ability to correctly classify light curves. This data-driven evidence strongly supports the conclusion that artificial intelligence can substantially aid human efforts in the challenging quest for exoplanets!

