Speaker
Description
We present a novel framework for cosmological model selection based on Neural Networks trained directly on simulated Cosmic Microwave Background (CMB) temperature and polarisation maps, as well as summary statistics.
A central component of this work is SkyExplain, an open-source framework designed to bring interpretability to machine learning applications on cosmological data. SkyExplain enables the mapping of feature importance directly onto the sky, bridging the gap between black-box models and physically meaningful insights.
We demonstrate the capabilities of SkyExplain in enhancing sensitivity to subtle signatures of primordial physics beyond the standard ΛCDM paradigm using CMB data. To this end, we generate Planck-like simulations and develop a hybrid architecture that combines principal component analysis with neural networks, optimised for classification tasks. We further interpret the model’s predictions using Shapley Additive exPlanations (SHAP), identifying the regions of the sky and angular scales that drive the distinction between ΛCDM and feature models.
The framework is designed to be easily extensible to a broad range of cosmological observables, from large-scale structure summary statistics to the identification of systematics in weak lensing shear maps. More broadly, our goal is to establish an open, community-driven hub where researchers can experiment with, adapt, and extend these tools for their own analyses.
We will present results on classification performance, calibration, and interpretability, alongside an Open Science initiative that provides public access to the full pipeline—from simulations to trained models and interpretability tools.
| Other topic / keywords: | cosmology, large-scale structure, cosmic microwave background |
|---|