A new method checks how models read spectra of HD 209458 b and HAT-P-26 b

A new approach helps explain how computer models infer exoplanet atmospheres from light. Tested on observations of HD 209458 b and HAT-P-26 b, it accounts for links between nearby wavelengths that simpler tests overlook. The results show which parts of a spectrum influence predictions and reveal that a neural network learned a known response reversal in the carbon dioxide signal.

Star system: HD 209458 | Distance from the Sun: 159 light years

Close

A new method checks how models read spectra of HD 209458 b and HAT-P-26 b

Star system: HD 209458 | Distance from the Sun: 159 light years

A new method can show which parts of an exoplanet’s spectrum influence a computer model’s predictions. This matters because machine-learning tools can quickly estimate atmospheric properties, but their internal reasoning is often difficult to inspect. The approach accounts for correlations between wavelengths, which can make simple tests of a model’s response misleading.

The method, called Gaussian perturbation, tests a model by making coordinated changes to spectral data based on the known links between wavelengths. Researchers compared it with a simpler approach that changes one wavelength bin at a time. They first used TauREx-III, a program that estimates atmospheric properties from spectra, as a stand-in for a black-box model. The tests used Hubble observations of HD 209458 b and James Webb Space Telescope observations of HAT-P-26 b. They then applied the method to a neural network that retrieves atmospheric properties for HAT-P-26 b.

For the gases that the models could constrain, Gaussian perturbation identified spectral regions with the expected importance. At the center of the carbon dioxide band in HAT-P-26 b, a degeneracy reverses the model’s response. A degeneracy occurs when different combinations of atmospheric properties can produce similar signals. The neural network showed the same reversal, indicating that it had learned this behavior. The method also identifies which parts of a spectrum a retrieval relies on, information that a model’s underlying physical equations do not provide on their own.

Key findings
Gaussian perturbation accounts for correlations between wavelengths when explaining regression models.
Tests used Hubble observations of HD 209458 b and James Webb observations of HAT-P-26 b.
For constrained gases, the method’s measure of spectral importance matched expectations.
At the center of HAT-P-26 b’s carbon dioxide band, a degeneracy reverses the response.
A neural network retrieval of HAT-P-26 b showed the same reversal.

Scientific study

Gaussian perturbation interpretability for regression techniques to understand retrieval black-boxes (external link)