Supernova Cosmology Inference – SCInf

The standard model of cosmology is confronted to two serious tensions: the current expansion rate is measured 5s faster than expectations given early cosmic age data (e.g. CMB) ; the dark energy equation of state parameter deviates at the 4s level from what is expected for a simple cosmological constant in Einstein’s equation. Cosmological parameters inference derived from Type Ia Supernovae (SN Ia) distances are at the core of both tensions. The question hence raises to know if these deviations are signs of new fundamental physics or unaccounted for systematics in Supernovae cosmological inference. With this ANR, we propose two new inference pipelines to tackle a key left-over questions: the interplay between selection effects and astrophysical evolution, since SN Ia physic is still largely unknown, but knowledge of SN Ia parameter properties are needed to account for selection biases. The two inference pipelines are closely connected in there development but differ in their concept: one is a multi-thousands free-parameters likelihood method, the other is a simulation based inference method that bypasses the need to explicitly write the likelihood by training a neural-network on realistic simulations. Both developments are enabled by recent machine learning libraries (jax) that offer GPU acceleration and automated analytical derivative. Both pipelines require realistic simulations. Developing them in parallel hence enables us to improve both technics while demonstrating which approach is the most suited for the future of SN cosmology. This research is made possible thanks to access to new datasets (ZTF at low-redshift ; HSC&SNLS5 at high-redshift) that have excellent calibration and simple selection effects. By the end of this ANR we will have measured cosmological parameters based on new inference methods and using a dataset of O(3000) SN Ia, none of which has been used by current cosmological analyses. We will then confirm (or not) current tensions.