2026 workshop program
An AI super-resolution field emulator for cosmological hydrodynamics: the Lyman-α forest
The work
Abstract
High-resolution cosmological simulations that include gas physics are essential for modeling the intergalactic medium, but they are computationally expensive and difficult to scale to large volumes. We present a deep learning framework that generates high-resolution baryonic fields from low-resolution simulations with high accuracy and dramatically reduced cost.
Our approach uses a two-stage model that first reconstructs small-scale structures from a low-resolution simulation and then refines them using high-resolution initial conditions. We apply this method at redshift z=3 and demonstrate that it accurately recovers key physical fields—including density, temperature, velocity, and optical depth—with sub-percent errors.
In addition, the model reproduces key observable statistics relevant to the Lyman-α forest, achieving ∼1% accuracy in the large-scale flux power spectrum and less than 10% error in the flux probability distribution function. Importantly, our method provides a computational speedup of approximately 260× compared to traditional hydrodynamical simulations at the same resolution.
This work demonstrates that AI-based emulation can serve as a fast and accurate alternative to full simulations, enabling the generation of large, high-resolution datasets for next-generation cosmological surveys.