How do adaptive mesh codes like sledgehamr handle singularities or extreme density gradients?

# How Adaptive Mesh Refinement Codes Handle Singularities and Extreme Gradients

Adaptive mesh refinement (AMR) codes like sledgehamr tackle the challenge of singularities and steep density gradients by dynamically concentrating computational resolution where it matters most—a strategy that's essential because simulating the entire domain at fine resolution would be prohibitively expensive.

The Core Strategy: Mesh Adaptation

Sledgehamr uses AMR built on the AMReX framework [AG-2024.04-069] to handle "large scale separations"—situations where you need fine resolution in some regions (near singularities or sharp gradients) but coarse resolution elsewhere. The code automatically refines the mesh in regions of interest and coarsens it where the solution is smooth, dramatically reducing memory and compute requirements compared to a uniform fine grid.

Handling Discontinuities and Shocks

When singularities manifest as shocks or discontinuities, pure high-order methods (like standard discontinuous Galerkin schemes) produce spurious oscillations. Several recent codes have adopted a hybrid approach: they use subcell finite volume limiters to suppress these oscillations while preserving high-order accuracy in smooth regions [AG-2024.06-410]. Think of this as a local "shock detector" that switches to a more conservative, lower-order method only where sharp features appear, then reverts to high-order elsewhere.

A refinement of this idea is the oscillation-eliminating discontinuous Galerkin (OEDG) scheme [AG-2024.10-129, AG-2025.12-072], which uses a linear damping equation to suppress spurious wiggles near discontinuities without requiring expensive characteristic decomposition. This matters because relativistic fluids (like neutron star material) must satisfy strict physical constraints—density and pressure must remain positive, velocities must stay subluminal—and oscillations can violate these constraints, breaking the simulation.

Managing Extreme Grid Stretching

A particularly tough problem arises in numerical relativity: when simulating black hole binary systems, boundary conditions must be imposed very far from the holes (to avoid artificial reflections), creating grids stretched by factors of a billion or more. Discontinuous Galerkin methods traditionally struggle here, but a new primal DG scheme [AG-2025.02-180] can handle stretching up to $\sim 10^9$ while maintaining exponential convergence. The key is careful discretization that doesn't break down on deformed elements.

Hybrid Approaches for Scalability

A newer strategy combines discontinuous Galerkin methods (which parallelize well) with compact finite volume methods [AG-2025.02-506]. By assigning DG to smooth-field regions and finite volume to shock-prone regions, this hybrid reduces communication overhead during AMR: each mesh element only talks to its six nearest neighbors (in 3D) rather than 26, dramatically improving scalability on supercomputers while making AMR implementation simpler.

Practical Trade-offs

Different discretizations suit different problems. For black hole mergers, vertex-centered grids are about 20% faster, while for neutron star mergers with magnetic fields, cell-centered discretization performs better [AG-2024.06-236]. This reflects a deeper principle: AMR effectiveness depends on matching the numerical method to the problem's dominant features.

---

The provided preprints don't directly discuss how sledgehamr itself selects or tapers refinement regions for scalar field singularities, though the framework clearly inherits standard AMReX refinement criteria.