Accelerating Transmission Spectroscopy of Exoplanets for Biosignature Detection Using Machine Learning

Overview

Detecting molecular species in exoplanet atmospheres through transmission spectroscopy is essential for exoplanet characterization, but traditional atmospheric retrieval methods are computationally intensive and difficult to scale to the large datasets expected from upcoming missions. The proposed solution is a machine learning pipeline using a dual-region Convolutional Neural Network (CNN) to infer atmospheric composition directly from transmission spectra more efficiently. I focused on detecting three molecules: H2O, CO2, and CH4. I created a hybrid dataset by combining synthetic spectra generated in Python containing molecule-specific absorption features with added Gaussian noise and spectral clutter to mimic real data, and published transmission spectra from the NASA Exoplanet Archive. All spectra were represented as transit depth ((Rp/Rs)^2) plotted with respect to wavelength, then standardized and formatted as inputs for the CNN. To reduce false positives from overlapping features and other molecules, the dual-region CNN analyzes each molecule independently and extracts features from two molecule-specific wavelength bands (R1 and R2), so a detection is supported by absorption in both regions rather than a single clustered peak. The combined dataset was split into 80% training and 20% testing. Model predictions on the held-out test set were compared to published results. Results show that the model can reliably detect H2O, CO2, and CH4, achieving >90% region-level accuracy across molecules, with many predictions made at high confidence (p > 0.65). A one-sample proportion (binomial) test with null hypothesis p̂ = 0.5 produced p-values<<0.01 for each molecule-specific model, indicating performance significantly better than chance.