Abstract

Ongoing and future cosmology experiments are poised to tackle some of the most profound questions in fundamental physics. These questions include explaining the nature of dark energy and dark matter, and testing particle physics and the theory of gravity on the largest observable scales in the universe. However, interpreting the results of these experiments presents a formidable challenge: deducing the underlying physics from acquired observational data, or solving an inverse problem. To rise to this challenge in the context of Lyman-alpha forest and intergalactic medium, we have developed the Nyx code, demonstrating remarkable efficiency and scalability. Yet, despite such computational capability, creating “virtual universes” remains extremely expensive, especially when high fidelity is required. Over the years, we have therefore experimented with combining simulations or physics-driven models with ML-based data-driven models, offering an efficient, albeit approximate, alternative to direct simulations. In this talk, I will briefly review different approaches we have tried and focus on our recent work aimed at improving the accuracy of low-resolution hydrodynamical simulations. This advancement allows us, for the first time, to produce simulations in Gpc boxes while maintaining accuracy of Lyman-alpha forest predictions comparable to high-resolution hydrodynamical simulations.