from abc import ABC, abstractmethod
import jax
import jax.numpy as jnp
import numpy as np
from fdtdx.colors import XKCD_LIGHT_GREY, Color
from fdtdx.core.jax.pytrees import autoinit, field, frozen_field
from fdtdx.materials import Material
from fdtdx.objects.object import OrderableObject
@autoinit
class StaticMultiMaterialObject(OrderableObject, ABC):
#: the static material
materials: dict[str, Material] = field()
#: the color of the material
color: Color | None = frozen_field(default=XKCD_LIGHT_GREY)
#: Enable sub-pixel (sub-cell) dielectric smoothing for this object. When ``True`` the assembler
#: replaces the binary voxel occupancy with an analytic fill-fraction and builds a smoothed,
#: anisotropic (full 3x3 tensor) effective permittivity at interface cells following Farjadpour et
#: al. (Meep): arithmetic mean of ``eps`` for the field components tangential to the interface and
#: harmonic mean of ``eps`` for the component normal to it. This removes the first-order staircasing
#: error of the Yee grid at strong dielectric jumps (2nd-order accuracy). Forces the whole
#: simulation to allocate an anisotropic permittivity tensor (3-component diagonal by default, or a
#: full 9-component tensor when ``subpixel_full_tensor`` is set). Requires the object to provide a
#: fractional ``get_fill_fraction_for_shape`` (the default falls back to the binary mask, which
#: still yields a valid but only cell-wide normal). See issue #373.
subpixel_smoothing: bool = frozen_field(default=False)
#: Selects the smoothing tensor representation when ``subpixel_smoothing`` is on. ``False`` (default)
#: keeps only the DIAGONAL of the Farjadpour tensor (``eps_ii = eps_bar - (eps_bar - eps_h)*n_i**2``),
#: allocating a cheap 3-component array that runs on the elementwise Yee update. This is EXACT for
#: axis-aligned interfaces (their normal lies on one axis, so the off-diagonal terms vanish) and is the
#: recommended production path for Manhattan geometries. ``True`` allocates the full 9-component tensor
#: (keeps the off-diagonal ``-(eps_bar - eps_h)*n_i*n_j`` terms), which is more accurate for tilted
#: interfaces (slanted sidewalls, diagonal edges) but ~3x heavier per step and forces the anisotropic
#: update kernel. Ignored when ``subpixel_smoothing`` is False.
subpixel_full_tensor: bool = frozen_field(default=False)
@abstractmethod
def get_voxel_mask_for_shape(self) -> jax.Array:
"""Get a binary mask of the objects shape. Everything voxel not in the mask, will not be updated by
this object. For example, can be used to approximate a round shape.
The mask is calculated in device voxel size, not in simulation voxels.
Returns:
jax.Array: Binary mask representing the voxels occupied by the object
"""
raise NotImplementedError()
@abstractmethod
def get_material_mapping(
self,
) -> jax.Array:
"""Returns an array, which represents the material index at every voxel. Specifically, it returns the
index of the ordered material list.
Returns:
jax.Array: Index array
"""
raise NotImplementedError()
def get_fill_fraction_for_shape(self) -> jax.Array:
"""Return the per-cell fill fraction of the object's material, in ``[0, 1]``.
This is the sub-pixel generalisation of :meth:`get_voxel_mask_for_shape`: interior cells return
``1.0``, exterior cells ``0.0`` and interface cells the fraction of the cell volume covered by
the object. The default implementation falls back to the binary mask cast to float, so a subclass
that does not compute a genuine fill fraction still behaves correctly (albeit without the
sub-pixel accuracy gain). Subclasses that can rasterise fractionally should override this.
Returns:
jax.Array: Float array of shape ``self.grid_shape`` with values in ``[0, 1]``.
"""
return self.get_voxel_mask_for_shape().astype(float)
def get_interface_normal_for_shape(self) -> jax.Array:
"""Return a per-cell unit interface normal derived from the fill-fraction gradient.
The normal is ``n = -grad(fill) / |grad(fill)|`` (the sign is irrelevant downstream because only
the symmetric outer product ``n ⊗ n`` is used). The gradient is taken with the object's physical
cell pitch on each axis, so the direction is geometrically correct on anisotropic grids. Cells
away from an interface (``|grad(fill)| ~ 0``) get a zero normal, which makes the smoothed tensor
collapse back to the isotropic bulk value. Computed in NumPy at initialisation (static geometry,
not a traced quantity).
Returns:
jax.Array: Float array of shape ``(3, *self.grid_shape)`` with the per-cell unit normal.
"""
frac = np.asarray(self.get_fill_fraction_for_shape(), dtype=float)
grid_shape = self.grid_shape
real_shape = self.real_shape
# Average physical pitch per axis (exact on uniform grids, a good local approximation on
# quasi-uniform grids; only the gradient *direction* matters after normalisation).
pitch = [(real_shape[i] / grid_shape[i]) if grid_shape[i] > 0 else 1.0 for i in range(3)]
grad = np.zeros((3, *frac.shape), dtype=float)
for ax in range(3):
if frac.shape[ax] > 1 and pitch[ax] > 0:
grad[ax] = np.gradient(frac, pitch[ax], axis=ax)
# n points from high fill (material) to low fill (background); sign is immaterial for n⊗n.
grad = -grad
norm = np.sqrt(np.sum(grad**2, axis=0))
safe = norm > 1e-12
normal = np.zeros_like(grad)
for ax in range(3):
normal[ax] = np.where(safe, grad[ax] / np.where(safe, norm, 1.0), 0.0)
return jnp.asarray(normal)
[docs]
@autoinit
class SimulationVolume(UniformMaterialObject):
"""Background material for the entire simulation volume.
Defines the default material properties for the simulation background.
Usually represents air/vacuum with εᵣ=1.0 and μᵣ=1.0.
"""
#: an integer values of the placement order
placement_order: int = frozen_field(default=-1000)
#: the static material
material: Material = field(
default=Material(
permittivity=(1.0, 1.0, 1.0),
permeability=(1.0, 1.0, 1.0),
),
)