Optimising Cosmological Emulators: Enhanced Training Data Sampling for Efficient Inference
Name of applicant
Andreas Nygaard
Institution
Aarhus University
Amount
DKK 1,437,275
Year
2024
Type of grant
Internationalisation Fellowships
What?
Cosmological inference aims to determine key properties of the Universe, like its expansion rate, using models based on physical laws. These models are computationally expensive, especially for complex N-body simulations, which track the movement of matter. Machine-learning models with the ability to emulate the results of these simulations can help reduce this computational burden significantly.
Why?
Understanding cosmology requires millions of model evaluations to compare theory with observations, but running full simulations is time-consuming and costly. Emulators allow researchers to achieve accurate results more efficiently, making large-scale projects, like mapping the Universe's structure, feasible within practical time limits.
How?
Emulators are trained on a small set of full simulations using clever data selection methods. Instead of running expensive simulations repeatedly, the emulator learns to predict outcomes for new scenarios in milliseconds. For high-cost models like N-body simulations, this drastically reduces computation time, enabling quicker cosmological insights.