Speaker
Description
Precision cosmology increasingly relies on repeated evaluations of computationally expensive observables, such as Cosmic Microwave Background (CMB) anisotropy spectra and large-scale structure statistics, posing a significant bottleneck for parameter inference and model comparison. Emulation techniques have emerged as a powerful solution, enabling fast and accurate interpolation of these observables across parameter space. I will begin by reviewing recent progress in observable emulation, highlighting their performance, limitations, and impact on modern cosmological analyses.
I will then present CLiENT (Cosmological Likelihood Emulator using Neural Networks with TensorFlow), a method that bypasses observable prediction entirely by directly emulating the likelihood function of a dataset given cosmological parameters. This approach provides a flexible and fully differentiable surrogate for the likelihood, enabling efficient gradient-based inference methods.
Using fewer than $\sim 2\times10^4$ training evaluations, the likelihood emulator achieves high fidelity, recovering posterior constraints to within $0.1\sigma$ of the true likelihood and maintaining pointwise accuracy at the level of $\Delta\chi^2\leq 0.5$ across relevant regions of parameter space. I will demonstrate the robustness and versatility of this approach, including applications to extended cosmological models.
These results position likelihood emulation as a powerful and complementary alternative to traditional observable-based approaches, with clear advantages for fast, flexible, and differentiable cosmological inference.