Entropy-Stable and Physical-Constraint-Preserving DGSEM for Symmetry-Reduced General-Relativistic Hydrodynamics on Stationary Spacetimes

What the math gives to ML

The paper separates geometry from local physical variables by evolving a weighted state and using a geometry-weighted mean as the admissible anchor. This suggests a transferable postprocessing layer for mesh and graph neural operators: preserve a weighted integral while contracting predicted nodal states toward a feasible anchor. Convexity makes the correction inexpensive and robust, avoiding a general nonlinear constrained optimization at every timestep. The strongest use case is autoregressive neural PDE solvers, where small positivity violations otherwise accumulate into unstable rollouts.