Parallel FDTD Simulation: Yee Grid and GPU Acceleration — enabling Maxwell's equations on structured grids

Finite-Difference Time-Domain (FDTD) solves Maxwell's equations on structured grids via explicit time-stepping. The Yee grid staggered arrangement (electric field at cell edges, magnetic field at cell faces) naturally implements curl operators via finite differences, avoiding numerical instabilities that plague collocated grids.

Yee Grid and Discretization

Time-stepping alternates E-field and H-field updates via curl operations: H_update ∝ ∇ × E, E_update ∝ ∇ × H. Courant-Friedrichs-Lewy (CFL) condition constrains timestep: Δt ≤ 1 / (c√(1/Δx² + 1/Δy² + 1/Δz²)). Violation causes numerical instability. This explicit scheme requires no matrix solve, enabling straightforward parallelization via stencil computation: each grid point independently updates using neighbors.

Ghost Cell Exchange and Domain Decomposition

Stencil kernels access neighboring grid points, requiring ghost cell exchange at domain boundaries. 3D FDTD decomposes spatial domain into rectangular tiles per MPI rank. At each timestep: compute interior points independently, exchange boundary planes with neighbors, update boundary points using received data. Overlapping communication and computation hides MPI latency: initiate ghost cell sends while computing interior stencils.

GPU FDTD Optimization

FDTD maps naturally to GPU: each thread updates one grid point (embarrassingly parallel). Shared memory caching of ghost values improves bandwidth utilization by 3-4x versus global memory access. Memory coalescing requires careful array layout: store fields in Fortran order (F-contiguous) to ensure adjacent threads access sequential memory addresses. Register usage per thread limits occupancy and register spill to local memory.

PML Absorbing Boundary Conditions

Perfectly Matched Layer (PML) surrounds the computational domain, absorbing outgoing waves via intermediate auxiliary variables that track field derivatives. PML updates follow the same stencil structure, doubling computational volume (outer PML region) but eliminating reflection artifacts. Parameter grading in PML optimizes absorption over frequency range.

Tools and Applications

MEEP (MIT Electromagnetic Equation Propagation) provides parallel FDTD with CUDA and MPI support. Photonics simulations (waveguides, cavities, metamaterials) and antenna designs (radiation patterns) exploit full-wave FDTD accuracy.