Academia Sinica Logo

Bayesian Surrogate Models: from Cosmological Simulations to Quasar Spectroscopic Pipeline

Abstract

In this talk, I will discuss the applications of Bayesian surrogate models in two areas of astrophysics. Firstly, I'll demonstrate how we utilize the multi-fidelity multi-scale emulation technique to create a surrogate model (emulator) from cosmological simulations with varying qualities. This approach effectively combines different resolutions and dynamic ranges of simulations, enabling accurate emulation of N-Body simulation codes for future Euclid and Roman galaxy surveys. Additionally, I'll present an example using PRIYA simulations, a galaxy formation suite based on the ASTRID simulation model with varied astrophysical feedback parameters. We apply this model to the eBOSS Lya flux 1D power spectrum and produce new constraints on cosmology. In the second part, I'll discuss how we construct a data-driven surrogate model for quasar emission, employing Bayesian model selection to probabilistically detect damped Lyman alpha absorbers and metal absorbers in SDSS quasar spectra. Our model encodes the full covariance of quasar emissions and the redshift evolution of optical depth, allowing us to provide probabilistic constraints even for low SNR spectra. This technique has been adopted by DESI for producing a damped Lyman alpha catalog, significantly improving the accuracy of the Lya 1D flux power spectrum.