Neural Modelling of Astrophysical Gas

Abstract

We investigate the abilities of neural models to model astrophysical gases, addressing limitations in traditional numerical solvers such as energy bottlenecks, resolution effects, grid anisotropy effects, and strict time-stepping constraints. In particular, we focus on simulations of turbulent flows and thermochemical networks in the interstellar medium (ISM). We employ the chaotic Kuramoto-Sivashinsky (KS) equation as a one-dimensional testbed for neural network architectures commonly employed in simulations of threedimensional hydrodynamical turbulence. We benchmark a wide range of state-ofthe-art neural architectures. Our experiments demonstrate that hierarchical context aggregation, residual connections, and group equivariance play a critical role in capturing faithful dynamics and spectral properties of turbulent flows. Additionally, we model the thermochemical evolution of astrophysical gas using a residual neural network (ResNet) trained on data generated by the CHIMES code. The network predicts the evolution of chemical abundances and thermal states across a range of densities, temperatures, and metallicities, effectively integrating over large timesteps to mitigate the stiffness issues of conventional solvers. Our findings suggest that modern deep learning methods may provide a viable and efficient alternative to traditional numerical solvers for astrophysical simulations