Explaining Neural Networks on the Sky: Machine Learning Interpretability for Cosmic Microwave Background Maps

We present a framework for cosmological model selection using Neural Networks (NNs) trained directly on simulated Cosmic Microwave Background (CMB) temperature and polarisation maps. By operating at the map level rather than on compressed angular power spectra, our approach retains the full spatial information of temperature and polarisation anisotropies, enabling the identification of subtle signatures of primordial features beyond the standard $\Lambda$CDM model. We describe the generation of Planck-like CMB maps, and the hybrid architecture that combines principal component analysis and neural networks, optimised for classification tasks. To understand how the classifier reaches its decisions, we apply Shapley Additive exPlanations (SHAP) as a post-hoc interpretability tool, identifying which regions of the sky and which scales contribute most to the distinction between $\Lambda$CDM and feature models. This work serves as a follow-up to previous analyses at the level of summary statistics and as a proof-of-concept for using interpretable machine learning to uncover higher-order information in CMB data, with the potential to enhance the detection of nontrivial inflationary signals and improve cosmological model discrimination. Results for model classification performance, calibration, and interpretability are presented as a placeholder for the full analysis. In addition, we introduce the Open Science project, providing public access to the full pipeline for simulation, training, and interpretability of CMB map-based neural networks.