Changelog¶

Unreleased¶

• Graded grids: a StackDevice runs on an fdtdx RectilinearGrid whose cells differ along any axis (every averaging). Fill fractions stay exact, each field component averages over the dual cells between cell centers along the axes it is staggered on, and averaging="subpixel_full" solves a junction voxel's cell problem on its own box (bf.subpixel.periodic_cell_tensor(..., sides=...)). Where the x and y cells differ, a voxel is painted as a rectangle with Euclidean normals and round corners; on graded x or y every column on its own (half-planes, corner model, sub-samples and junctions), with roughness drawn on an even lattice and interpolated to the columns, and finite-difference normals from the distances between cell centers. bf.graded_edges places graded edges (fine regions, a growth limit, interfaces on edges, uniform cells under the PML), StackDevice.rectilinear_grid builds the grid over the box, and the create_*_at methods take grid= to give port planes whole cells there. Uniform grids are unchanged, bit for bit.
• The viewer adopts results recorded on non-uniform grids (detectors at their physical extent, from the grid's edges) and draws their field planes resampled onto an even grid at each graded axis's finest cell.
• averaging="subpixel_full": a junction sub-cell whose two largest face directions weigh nearly the same (a polygon corner in it) blends its laminate toward a direction-free one instead of switching it between the walls' normals at round-off, so the written tensor is continuous there (a corner on a cell's corner made it jump by about 2e-3, and finite differences did not converge). Other sub-cells keep their laminate exactly; the closure's accuracy against the finite-element reference (benchmarks/subpixel/validate_closure.py) is unchanged.
• place_on_grid warns when a level that from_layer_stack left out (outside the box by more than level_margin and a 0.1 um slack) is sampled by a coarse cell at the box's face; build the device with a larger level_margin then.

0.2.0¶

• Inverse design guide and example (examples/guide/inverse_design.py): the width on each side of an air-clad silicon nitride bend, at 70 stations, optimized for TE0 transmission through the simulation (56 % to 99.5 %; the uniform bend converts most of the rest to TE1).
• bf.jit: jax.jit with an XLA option (simplify-conditional disabled) for gradients through fdtdx's checkpointed time loop, which plain jax.jit compiles into a backward loop that computes a time step's VJP on every iteration. The inverse-design example compiles its whole optimization step with it.
• Dependencies: gdsfactoryx 9.51.0.post2 or newer, whose geometry traces under jax.jit, so a whole objective that builds a component can be compiled, and whose GDS output, reference names and cross-section names match upstream gdsfactory's.
• Installation docs are backend-neutral: bifrost runs on any JAX backend; set JAX up with the JAX installation guide.
• The viewer design spec moved out of the documentation site to design/viewer-spec.md in the repository.

0.1.0¶

First release on PyPI, as bifrost-photonics (the import is still import bifrost as bf); the fdtdx stack and viewer moved here from gdsfactoryx.

• bf.StackDevice: a gdsfactory component and its LayerStack voxelized into one differentiable fdtdx object, with sources and monitors placed on ports: sidewalls, derived layers, exact corners, subpixel and full-tensor averaging, roughness, dispersive and anisotropic materials, diagnostics and timing.
• StackDevice(dtype=...): the whole simulation in float32 (default) or float64.
• bf.view: the 3D viewer widget for setting up, running and inspecting a simulation, with a precomputed live demo.
• Examples: the waveguide bend tutorial, and the elevator and directional couplers compared with MEEP.
• Dependencies: fdtdx (required, no extra), gdsfactoryx from PyPI, tidy3d 2.12 or newer.