Description
Recent spectral observatories stand to revolutionise our ability to study exoplanets on a larger population scale than ever before. Analysing this data requires extracting information about the planetary atmosphere from the spectra. For anything outside a small number of targets this is very computationally resource intensive, which is a large barrier to entry as we move into larger scale planetary surveys.
The use of ML has been proven as a powerful tool in tackling this, reducing computational resources required. However, the scale of these models means this advantage comes at the cost of understanding.
We go beyond existing approaches, presenting a novel method of interpretability based on physically motivated forward modelling, bridging the gap between ML and traditional exoplanet retrieval approaches.
We trained a range of neural networks to predict the atmospheric abundances of molecules across 40,000 simulated spectra for the Ariel space telescope, then compare a selection of existing techniques for interpreting predictions, including SHAP (SHapley Additive exPlanations). Based on this analysis we propose a novel application of the perturbation sensitivity technique for interpreting ML predictions, which is shown to be more aligned with physical models.
This method has potential for use outside of space-based exoplanetary data, and we believe the opportunity to share it here would help unlock barriers to entry in the use of ML for astrophysical spectral analysis across fields.
