Answers are grounded in this paper and cite their sources.
The potential of deep learning-based image-to-image translations has attracted significant attention as a fast, approximate alternative to cosmological simulations, which would be particularly useful in various contexts, including covariance studies, investigations of systematics, and cosmological parameter inference. We model the evolution of structure formation using U-Net, a widely employed convolutional image translation framework with previous applications in cosmology. To study aspects of learning-based cosmological mappings in a controlled, illustrative setting, we use two approaches for generating suitable cosmological matter fields as datasets: a simple analytical prescription provided by the Zel’dovich approximation, and a standard N-body method using Particle-Mesh simulations. We apply established metrics such as matter power spectra and bispectra, topological indicators, and statistical independence of density fields and study their reproducibility and convergence properties. The U-Net approach performs well only for some of these physical metrics, and accuracy is worse at more nonlinear scales, where the dynamic range in density is larger. Motivated by field-level emulator studies utilizing Eulerian-aware loss functions and beyond pixel-wise mean squared errors, we examine a density-weighted objective. In the PM tests ending at z = 4, where the evolution is only mildly nonlinear, using the Eulerian matter density evaluated at each particle position as a direct weight on that particle’s Lagrangian displacement error improves the U-Net results at smaller scales; this result does not establish comparable accuracy in the strongly nonlinear low-redshift regime.
No takes yet. Share an insight, caveat, or question.
Dong et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: