# Capability roadmap

The aim is not to clone Tidy3D's web platform. It is to approach the functionality and confidence of its FDTD workflows with an open, local, accelerator-native solver.

## Strong today

- JAX/GPU Yee stepping with reverse-mode differentiation;
- dielectric, conductive, dispersive, anisotropic, and complex tensor media;
- PML, PEC, PMC, periodic and Bloch boundaries;
- plane, Gaussian, dipole, mode, and TFSF sources;
- field, energy, phasor, Poynting, mode-overlap, and projection detectors;
- automatic grid/boundary setup with a voxel budget, exact planar-interface
  alignment, and physical-thickness PML sizing;
- primitive geometry, extruded polygons, GDS-layer helpers, and subpixel smoothing;
- straight-waveguide, splitter, Bragg, and ring-bus forward paths;
- reusable filtering/projection, D4 symmetry, erosion/dilation, binary
  validation, bounded optimization, and reverse-tape memory planning;
- exact notebook-derived inverse-design adapters with fully binary validation;
  the D4 S-matrix crossing is independently confirmed within 10% but remains
  below 95% parity, while the other tested devices remain explicit failures
  rather than local-score passes;

## Near-term priorities

1. **Spectral ergonomics:** standard NumPy/xarray result containers with units, frequency axes, reference-plane metadata, and normalization provenance.
2. **Mode-port robustness:** reflected-port calibration, angled/bent guides, multi-port automation, and clearer propagation conventions.
3. **Reference repair:** regenerate the three quarantined Tidy3D goldens under converged, reproducible settings after explicit authorization.
4. **Grid refinement:** true local refinement beyond per-axis rectilinear spacing, especially around thin slabs and curved interfaces; recorder-memory planning and physical PML sizing are now automatic.
5. **Inverse-design geometry:** differentiable polygon/shape controls and curvature constraints for Autograd8-class bends, plus local refinement/checkpointing for Autograd25-scale full 3D tapes.
6. **Optimizer usability:** restartable state and process-corner orchestration on top of the present high-level objective loop, callbacks, continuation, fabrication constraints, and held-out binary validation.

## Later FDTD coverage

- astigmatic/focused and custom field/current sources;
- near-to-far and diffraction parity;
- tabulated optical constants and multi-pole fitting;
- resonator $Q$ extraction and robust termination;
- periodic/metasurface sweep helpers;
- complex boolean geometry and STL robustness;
- distributed-device examples and profiling.

## Explicitly outside this program

Heat and charge solvers, EME, GUI/web API replication, and unrelated multiphysics are not part of the current FDTD parity target. They should not dilute correctness work on Maxwell simulation and optimization.