Orateur
Description
The cosmic web, consisting of an intricate network of voids, walls, filaments and clusters, contains key information about cosmological parameters and the evolution of large-scale structures in the universe. In this context, very large ensembles of cosmological simulations are required to analyse upcoming galaxy survey data. To circumvent the associated computational cost, we aim at constructing a generative diffusion model to emulate the cosmic web while preserving its statistical properties. In this context, we explore how self-attention maps, which are learnt neural network features embedded within the diffusion model, can offer a new way to probe the underlying structure of the cosmic web. In addition to emulating high-fidelity simulations, our approach could also provide a quantitative assessment of the performance of generative models.