Interpretability attribution maps of neural networks trained on Cosmic Microwave Background data

This dataset contains SHAP (SHapley Additive exPlanations) attribution maps derived from a Neural Network (NN) trained to perform model selection on Cosmic Microwave Background (CMB) temperature and polarisation maps (T, Q, U), as part of the analysis in:
https://github.com/SkyExplain/SkyInterpret

The SHAP maps were computed using the partition explainer algorithm applied directly to the classified CMB images, and subsequently reprojected from Cartesian to full-sky HEALPix format to preserve the geometric structure of the CMB sky. They quantify the pixel-level contribution to the neural network's classification decision between the standard ΛCDM model and a model featuring an oscillatory template in the primordial power spectrum, parametrised as:

P_R(k) = P_{R,0}(k) [1 + A_lin sin(ω_lin k/k* + φ)]

where the feature amplitude A_lin and feature frequency ω_lin are varied across the dataset.

The input CMB maps used to train the NN, which correspond to these SHAP attributions, were simulated using CAMB and healpy with the official Planck galactic mask and Planck-like noise applied, and are available at: 10.5281/zenodo.19445834 

The trained neural network weights used to produce these SHAP values are available at: 10.5281/zenodo.19445671

The full simulation pipeline is available at: https://github.com/SkyExplain/SkySimulation
The model selection pipeline is available at: https://github.com/SkyExplain/SkyNeuralNets