ADR 0002: Spatial operators — FD vs spectral (direction)

Status

Proposed / in progress (implementation spans multiple releases)

Context

  • Spectral methods (FFT via HeFFTe) are the default backbone for shipped PFC-style models: linear operators in k-space, nonlinear terms in real space.

  • Spectral discretizations can be limiting for some boundary treatments, localized operator control, or formulations where global FFT assumptions are awkward.

  • Finite-difference stencils already integrate with the same decomposition and halo machinery (halo_exchange.md); kernel helpers live under pfc::field::fd (finite_difference.hpp).

Decision

Advance a unified abstraction for spatial operators (gradients, Laplacians, and related terms) so that, where supported, callers can select spectral versus FD evaluation consistently with domain decomposition and halo policy. Shipped spectral JSON apps may remain FFT-centric until wiring is complete; new and internal models can adopt the abstraction earlier.

Consequences

  • Documentation must describe both the current spectral-first reality and the FD roadmap (when_not_to_use_openpfc.md).

  • Performance and correctness tests must cover both paths where exposed.

  • refactoring_roadmap.md may track concrete milestones.

See also