Field-level weak lensing cosmology with < 100 simulations using multifidelity simulation-based inference

Research Question and Approach

Extracting cosmological information from weak lensing surveys requires moving beyond two-point statistics to capture non-Gaussian information at smaller, non-linear scales. However, this necessitates high-fidelity simulations that are computationally expensive to produce. The authors investigate whether multifidelity simulation-based inference (SBI) can mitigate this cost. They propose a transfer learning framework where neural density estimators and compression models are pre-trained on fast, low-fidelity log-normal simulations (generated using the GLASS package) and subsequently fine-tuned on a small set of high-fidelity N-body simulations (from the Gower Street suite).

The study demonstrates that this multifidelity approach enables an order-of-magnitude reduction in the number of high-fidelity simulations required for accurate inference. By leveraging the pre-trained models, the researchers show that between 60 and 100 high-fidelity simulations are sufficient to obtain well-calibrated cosmological posteriors. This method successfully incorporates realistic survey systematic effects, such as photometric redshift uncertainties and halo-mass-dependent intrinsic alignment models, while maintaining computational efficiency.

As Stage-IV cosmological surveys like Euclid approach, the ability to perform field-level inference—which exploits the full information content of the lensing field—is critical. The high computational cost of high-fidelity simulations has historically been a bottleneck for such analyses. This work provides a scalable path forward, enabling researchers to utilize state-of-the-art simulations without the prohibitive cost of generating thousands of high-fidelity realizations, thereby facilitating more precise constraints on dark energy and matter composition.