12:00–1:00 pm ERC 401
Host: Pato Gallardo
Ming-Feng Ho (UC Riverside) "Multi-fidelity emulation for matter power spectrum and Lyman alpha forest using Gaussian Processes"
Next-generation telescopes (e.g., Roman space telescope) will probe the small-scale structures of the Universe with high accuracy. Thus, expensive large-volume high-resolution simulations are required to extract cosmological information until non-linear scales, making Bayesian parameter inference impractical for future surveys. One solution is multi-fidelity emulation, which uses simulations of varying qualities to train an accurate emulator at a much lower cost. We present the first implementation of multi-fidelity emulation in cosmology, where we use machine learning to model summary statistics from different resolutions of cosmological simulations using a sequence of Gaussian processes. We show two use cases of our emulation framework: the matter power spectrum from N-body simulations and the Lya flux power spectrum from hydrodynamical simulations. Our proposed multi-fidelity emulator offers a new way to predict non-linear scales, making the emulation development substantially more practical for future cosmological inference problems.
A short bio: I am an astrophysics Ph.D. student and a NASA FINESST FI at UC Riverside, working with Simeon Bird on various projects in Bayesian statistics and machine learning. My research interests include intergalactic medium, numerical cosmology, black holes, and Bayesian calibration for computer codes (emulation). My past work includes using Gaussian processes to find strong Lya absorbers and estimate quasar redshifts. I am currently working on combining cosmological simulations from different qualities to build multi-fidelity emulators using Gaussian processes.
