Interpreting Atmospheric Neural Networks: Insights from Explainable AI

Abstract We demonstrate how explainable artificial intelligence (XAI) methods can increase confidence in neural network predictions for exoplanet atmospheric modelling. We analysed a convolutional U-Net trained to reconstruct disequilibrium steady states from equilibrium chemistry states. We evaluated how the model encodes and uses physically meaningful features by combining exploratory data analysis with XAI techniques. Exploratory data analysis revealed patterns in the data depending on initial parameters such as the carbon-to-oxygen ratio, metallicity, and temperature. We observed the discovered dependencies throughout the network, confirming that the network learns relevant patterns rather than spurious relationships. By tracking the most influential molecules, we show that the network dynamically adjusts its focus with changing initial parameters, demonstrating that it captures physically meaningful differences between atmospheres. Furthermore, the tests performed also revealed limitations in the network that had not been considered previously. This work highlights how interpretability tools clarify learned representations to improve scientific reliability and can inform architectural refinements due to the limitations found.