Astro Seminar Series - Marco Bonici - HYBRID

Marco Bonici is a postdoc at the Waterloo Centre for Astrophysics at the University of Waterloo. Originally from Genova, in 2017 he completed his Bachelors and Masters degree at Università degli Studi di Genova in Theoretical Physics. He completed his PhD in Physics and Astrophysics in 2021 at University of Genova, under the supervision of Dr. Carmelita Carbone. Marco is an active member of the Euclid Consortium, mainly working in the Galaxy Clustering working group and at the development of the Euclid official likelihood pipeline. Recently, he has moved his focus on the development of techniques to accelerate the analysis of present and forthcoming surveys, from surrogate models to gradient-based algorithms.

Title: Fast and differentiable emulation of cosmological observable

Abstract: Forthcoming LSS and CMB surveys will provide us with an unprecedented amount of data, that will shape our knowledge of the Universe. However, the analysis of these datasets poses a serious computational challenge, pushing standard methods to their limits. In order to face this kind of problem, several codes known as emulator have been developed by the cosmological community. An emulator is an approximate method that mimics the behavior of a computationally expensive function. Although creating an emulator is conceptually straightforward, there are several design choices (type and architecture of the surrogate model, preprocessing procedure, dimensional reduction technique, ...) which can have a dramatic impact on the final performance. In this talk I will present one of these applications, Capse.jl, a novel emulator that utilizes neural networks to predict Cosmic Microwave Background (CMB) temperature, polarization and lensing angular power spectra, which uses different preprocessing and dimensional-reduction techniques to similar works in the literature. The emulator computes predictions in just few microseconds with emulation errors below 0.1 sigma for all the scales relevant for the planned CMB-S4 survey. Capse.jl can also be trained in an hour's time on a CPU. As a test case, we use Capse.jl to analyze Planck 2018 data and ACT DR4 data. We obtain the same result as standard analysis methods with a computational efficiency 3 to 6 order of magnitude higher. We take advantage of the differentiability of our emulators to use gradients-based methods, such as Pathfinder, Hamiltonian Monte Carlo (HMC), and MicroCanonical Hamiltonian MonteCarlo (MCHMC), which speed up the convergence and increase sampling efficiency. Finally, I'll give an overview of similar ongoing applications in the cosmological context and how I am working to extend the developed framework.