Prize-Winning Science Fair Project by Maximilian Solomon ’30

Exoplanets, quite simply, are planets beyond our solar system. Discovering and learning more about them is crucial in understanding universe formation, composition, and habitability; id est, the search for extraterrestrial life. There are estimated to be one septillion stars in the universe (10^24 or 1,000,000,000,000,000,000,000,000 stars!) yet only 5,830 exoplanets have been found. NASA, ESA and other prominent global space agencies have spent billions of dollars on both land- and space-based telescopes to collect data for the detection of these exoplanets. However, when exoplanets are suggested by a telescope, they must be confirmed by at least two other telescopes’ data. Moreover, the actual analysis of this data is slow and inconsistent.

Maximilian’s project was to analyze exoplanet data with machine learning, a computer science technique of training a “model” to learn patterns (temporal or frequential) in statistical data to make accurate, logical predictions of an output value, in this case, whether there was an exoplanet represented by the data or not. Specifically, Maximilian used transit photometry data; this works on the principle of transits, where exoplanets orbiting around their host star cause a small, periodic dip in the flux, or brightness, of those stars. A star’s brightness is recorded over time on a graph called a light curve, and if there are consistent dips in flux, there are transit events, signifying a planet orbiting the star. Machine learning had previously been integrated to facilitate the difficult confirmation task but with low recall (sensitivity) and accuracy, often not appropriately addressing the data imbalance between stars with exoplanets and stars without exoplanets, or sufficiently enhancing transit patterns and obscuring noise.

Maximilian had two goals: to create a novel data augmentation methodology that could boost ML model performance, specifically in the metrics of recall and accuracy to facilitate the exoplanet confirmation task and evaluate the efficacy of the data pipeline on various ML models. Maximilian acquired 5,657 stars (99.3% non-exoplanet and 0.7% exoplanet) from a publicly available NASA dataset (preprocessing the data) and constructed 7 ML models (KNN, LR, RF, CNN with 8 layers, CNN with 6 layers, CNN-RNN hybrid, and MLP). He then developed the augmentation methodology. He converted the temporal flux data to frequency domain with Fast Fourier Transform, representing underlying transit patterns as low amplitude, high frequency components (making them more prominent). He used the Savitzky-Golay noise-reduction filter to smooth the frequency components representing noise, enhancing the transit components. He applied a normalization formula to the data for consistency, especially for the benefit of the deep learning models during training. To end, he applied a data synthesis methodology (synthetic minority oversampling technique, or SMOTE) that, instead of replicating data for the minority class, synthesized exoplanet star data based in between preexisting data points, boosting model generalization for real-world scenarios. He trained and tested all of the models on both the original and augmented data and found a significant increase in recall and accuracy of the augmented data results, proving the methodology successful. Moreover, he found that his highest scoring models, his CNNs, with 99.65% accuracy and 100% recall (perfect for the confirmation task), were higher than other professional studies evaluating ML on exoplanet detection with transit photometry data.

He hypothesizes this methodology, HELIOS (High-Accuracy, Exoplanet Detection via Light-curve Interpretation with Optimized Fourier Analysis and SMOTE Synthesis), can be used to find exoplanets with other temporal-based detection methods (like radial velocity, gravitational microlensing, astrometry, etc.) as well as detect USPs (ultra-short period planets) or rogue planets. He believes there is also potential for HELIOS in nonbinary classification, that is, detecting how many planets there are in a potentially multiplanetary system. Maximilian entered his project into the Middle School Science Fair this year and won first prize. He then advanced to the NCSEF and NCSAS science fairs. He won both competitions in first place at the regional level and is now advancing to state. He is currently applying his CNN-6 onto a publicly available website for the thriving NASA citizen-science community, producing hundreds of thousands of light curves every year. He hopes the model helps the community find an exoplanet quickly and accurately and wants to conduct further research on the methodology.