Oral Presentation

Towards field-level multi-probe analyses using differentiable hydrodynamics

Author(s): Tilman Troester (ETH Zurich)

Presenter: Tilman Troester (ETH Zurich)

Two approaches to extract more information from our data are multi-probe analyses and field-level inference. In the former, complementary observables are analysed jointly to mitigate systematics and break parameter degeneracies, while in the latter the data are analysed without resorting to summary statistics, thus avoiding loss of information.
So far, multi-probe analyses have largely relied on two-point summary statistics, while field-level inference has focussed on galaxy clustering and weak lensing.
The next logical step is therefore to combine these approaches and aim for field-level multi-probe analyses, in particular by including observations of gas to constrain and mitigate the effect of baryon feedback on weak lensing. This necessitates differentiable forward models, and in the case of baryonic probes, differentiable hydrodynamics.
In this talk I will present preliminary results of ongoing work on hydrox, a differentiable and GPU-accelerated smoothed-particle hydrodynamics code written in JAX.
Such a differentiable simulation code gives unprecedented control and flexibility over the sub-grid physics during data analysis, as it enables directly fitting the feedback processes to the data at the field level, such as weak lensing and tSZ observations, instead of relying on interpolating between summary statistics of pre-calibrated, discrete sets of simulations.