Lunch Talk: Ming-Feng Ho (UCR)
Next-generation telescopes like the Roman Space Telescope will provide unprecedented accuracy in probing small-scale structures of the Universe. However, extracting cosmological information from these surveys requires expensive high-resolution simulations spanning the high-dimensional parameter space, making Bayesian parameter inference impractical. We propose using multi-fidelity emulation to overcome this problem, where simulations of varying qualities are used to train an accurate emulator at a lower cost. Our implementation uses machine learning and Gaussian processes to model summary statistics from different resolutions of cosmological simulations. We demonstrate the effectiveness of our approach with two use cases: the matter power spectrum from N-body simulations and the Lya 1D flux power spectrum from hydrodynamical simulations. Our proposed multi-fidelity emulator offers a practical way to predict non-linear scales, making emulation development more feasible for future inference problems.
813 Santa Barbara Street
Pasadena, CA 91104
United States
![spectrum](/sites/default/files/styles/large/public/2023-03/spectra_simulation.png?itok=v3qTxxk1)