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

Researchers demonstrate that multifidelity simulation-based inference can extract cosmological information from weak lensing fields using fewer than 100 high-fidelity N-body simulations, achieving an order-of-magnitude reduction in computational cost. By pre-training neural models on fast, low-fidelity simulations and fine-tuning on expensive high-fidelity runs, the method enables field-level cosmological inference that captures substantially more information than traditional two-point statistics.

This research addresses a fundamental computational bottleneck in modern cosmology: extracting maximum information from observational data while minimizing the prohibitive cost of physics simulations. Traditional cosmological inference relies on summary statistics like power spectra, but field-level analysis captures richer information encoded in spatial structures. However, this requires physically realistic simulations that are computationally expensive, creating a tension between data richness and practical feasibility.

The multifidelity approach solves this by leveraging a hierarchy of simulation fidelities. Fast log-normal GLASS simulations serve as a training foundation for neural compression and inference models, while a small cohort of expensive N-body simulations fine-tunes these models for accuracy. This strategy mirrors successful approaches in engineering and machine learning, where surrogate models reduce reliance on costly high-fidelity evaluations.

For the broader cosmology and computational science communities, this work has significant implications. The demonstration that 60-100 high-fidelity simulations suffice for well-calibrated posteriors democratizes field-level inference, enabling smaller research groups and institutions with limited computational resources to conduct competitive cosmological analyses. The neural compression pipeline also hints at applications beyond weak lensing, potentially extending to other data-intensive observational domains requiring expensive simulations.

The methodology's success validates multifidelity machine learning as a practical tool for simulation-intensive science. Future developments might explore adaptive sampling strategies to further optimize simulation allocation, or extend the framework to joint analyses combining multiple observational probes, pushing the boundaries of what limited computational budgets can achieve in precision cosmology.

• →Multifidelity simulation-based inference reduces high-fidelity simulation requirements from thousands to 60-100 for field-level cosmological analysis
• →Neural compression of weak lensing fields captures substantially more cosmological information than traditional two-point summary statistics
• →Pre-training on fast, low-fidelity simulations followed by fine-tuning on expensive runs enables efficient transfer learning in cosmology
• →This approach democratizes field-level inference by making it computationally accessible to research groups with limited resources
• →The methodology demonstrates order-of-magnitude cost reduction while maintaining well-calibrated and informative cosmological posteriors