Transfer learning for multifidelity simulation-based inference in cosmology

TL;DR

This paper introduces a transfer learning approach that combines low- and high-fidelity simulations to perform efficient and accurate cosmological parameter inference, significantly reducing computational costs in simulation-based inference.

Contribution

The authors develop a multifidelity transfer learning method that decreases the need for expensive high-fidelity simulations in cosmology SBI by leveraging cheaper low-fidelity simulations.

Findings

Pre-training on low-fidelity simulations reduces high-fidelity simulation requirements by 8 to 15 times.

The method achieves accurate inference with substantially lower computational costs.

Demonstrated on dark matter density maps from the CAMELS dataset.

Abstract

Simulation-based inference (SBI) enables cosmological parameter estimation when closed-form likelihoods or models are unavailable. However, SBI relies on machine learning for neural compression and density estimation. This requires large training datasets which are prohibitively expensive for high-quality simulations. We overcome this limitation with multifidelity transfer learning, combining less expensive, lower-fidelity simulations with a limited number of high-fidelity simulations. We demonstrate our methodology on dark matter density maps from two separate simulation suites in the hydrodynamical CAMELS Multifield Dataset. Pre-training on dark-matter-only $N$-body simulations reduces the required number of high-fidelity hydrodynamical simulations by a factor between $8$ and $15$, depending on the model complexity, posterior dimensionality, and performance metrics used. By leveraging cheaper simulations, our approach enables performant and accurate inference on high-fidelity models while substantially reducing computational costs.