Efficient Surrogate Modeling for Fast Hydrodynamic Evolution of the Quark-Gluon Plasma

We propose a neural operator-based surrogate model for fast simulation of 2+1D and 3+1D quark-gluon plasma (QGP) hydrodynamic evolution. Accurate event-by-event simulations require solving computationally expensive PDEs, often taking hours per event. While neural operators enable efficient operator learning, existing approaches assume static domains and do not explicitly capture dissipative dynamics, evolving spatial support, or scale efficiently to spatiotemporal settings. We introduce a time-conditioned neural operator that jointly models field evolution, energy dissipation, and dynamic spatial support. Our approach combines a factorized spatial neural operator with a time-conditioned decoder that adaptively modulates latent features over proper time, along with a gated spatial masking module to capture the evolving support of the plasma fireball. Experiments on simulation data generated by the IP-Glasma+MUSIC framework show accurate predictions of key physical observables while achieving orders-of-magnitude speedup over traditional solvers. These results highlight the potential of our proposed model as a scalable and high-fidelity surrogate for QGP hydrodynamics.