Learning Dissipative Subgrid Physics in Relativistic Hydrodynamics with Neural Operators

Relativistic dissipative fluid dynamics play an important role in describing the collective flows in heavy-ion collisions and high-energy astrophysical flows, such as black hole accretion disks and binary neutron star mergers.Accurate simulations of relativistic hydrodynamics require resolving the small-scale turbulence to correctly capture the dissipative effects modeled as shear stress, bulk viscosity, and heat flux. These quantities regulate transport processes and mediate the stability of global flow dynamics, but directly resolving the small-scale dynamics relevant in a global simulation is computationally prohibitive.Existing sub-grid models remain expensive, as evaluating the full set of dissipative contributions requires substantial numerical effort. We present a neural-operator framework that surrogates microscale dissipative physics, typically handled by expensive implicit solvers, and injects their effects into large-scale relativistic simulations through effective macroscopic parameters. We estimate that the proposed framework has the potential to accelerate traditional global simulation of black hole accretion disk by at least four orders of magnitudes with proper microphysics feedback, enabling multi-scale simulations of relativistic flows that were previously intractable.