Likelihood-free inference with Dark Energy Survey Y3 data

Likelihood-free inference with Dark Energy Survey Y3 data

Project summary:

Using optimised “active learning’’ to efficiently sample the cosmological model space, we will generate a suite of simulated catalogues of mock Dark Energy Survey (DES) data. These mock data are the primary input (along with the actual observed data) to the DES likelihood-free inference pipeline. This pipeline uses state-of-the-art statistical methods to avoid many current assumptions that are known to be likely sources of error, and which can lead to incorrect cosmological inference or the creation of artificial model tensions. This project uses a new simulation-based inference approach creating a new frontier for DES to understand dark energy, dark matter, neutrinos, and the initial conditions of the Universe.

The result of this proposed research will significantly improve the robustness of new discoveries and of cosmological model inference from the DES Year 3 Weak Lensing analysis (one of the two main DES probes). The resulting methodology, code and simulation products will then be further incorporated in the final (Year 6) DES analysis in the following year (2022+). The DiRAC hardware provides the only machine available in the UK to carry out this full analysis.

Project Science Highlight:  We cannot currently share the cosmology “science” results (including inference for dark energy models) due to DES policy regarding unblinding at this stage. We can, however, share a visualisation of a typical simulation that has been run on a GPU as part of the suite generated in the project.

Image description: This is a visualisation of the output of the simulation – a dark matter map. This is a “view from Earth” looking out into the simulation over a certain radial distance. The structures you can see in the image are bright spots representing high density of matter and darker spots showing lower densities of matter (cosmic voids). Overall you can see the cosmic web.

The structures in this image are expected to change depending on the cosmological model. For example, if we have a different value for the Hubble parameter or a different equation-of-state for Dark Energy, these structures would appear different. The structures would also appear to be at different distances from us. 

By using the DiRAC GPUs, we are able to run multiple realisations of different Universes that can be compared to the observed dark matter map (from the Dark Energy Survey). In particular, as we can run fast independent simulations in parallel (with one simulation per GPU to avoid message passing overheads) we are able to sample many cosmological parameters. We can then use deep learning techniques and methods of “likelihood-free inference” to estimate our uncertainty in those cosmological parameters given the observed DES data – this has never been done  yet with a full galaxy survey.