Tutorial example for computing the adjoint gradient of a level set (#…

…2792) * tutorial example for computing the adjoint gradient of a level set * update figures and add results for S and P polarization * replace smoothing of the levelset using a signed-distance function with downsampling from a high-resolution grid * update figures for backpropagation calculation and Jacobian matrix * add figures showing simulation layout and plot of error in adjoint gradient vs resolution * more details on defining planewave using EigenmodeCoefficients object * add text describing simulation setup and results * Update doc/docs/Python_Tutorials/Adjoint_Solver.md * Update doc/docs/Python_Tutorials/Adjoint_Solver.md * Update doc/docs/Python_Tutorials/Adjoint_Solver.md * Update doc/docs/Python_Tutorials/Adjoint_Solver.md * Update doc/docs/Python_Tutorials/Adjoint_Solver.md * Update doc/docs/Python_Tutorials/Adjoint_Solver.md --------- Co-authored-by: Steven G. Johnson <[email protected]>
NanoComp · Jun 28, 2024 · 833e380 · 833e380
1 parent 88022e8
commit 833e380
Show file tree

Hide file tree

Showing 6 changed files with 732 additions and 0 deletions.
diff --git a/doc/docs/Python_Tutorials/Adjoint_Solver.md b/doc/docs/Python_Tutorials/Adjoint_Solver.md
@@ -804,6 +804,392 @@ if __name__ == "__main__":
         )
 ```
 
+Derivatives with Respect to Shape Parameters
+--------------------------------------------
+
+It is also possible to compute the derivative of the Meep outputs with respect to a geometric parameter via a [level-set](https://en.wikipedia.org/wiki/Level_set) formulation (an implicit-function representation of a material discontinuity) using the density-based adjoint solver. This is useful for applications involving [shape optimization](https://en.wikipedia.org/wiki/Shape_optimization) of explicitly parameterized geometries.
+
+As a demonstration, we will compute the derivative of the diffraction efficiency of the first transmitted order with $\mathcal{S}$ polarization (electric field out of the plane, i.e. $E_z$) of a 1D binary grating with respect to the grating height. The accuracy of the adjoint derivative is validated using a brute-force finite-difference approximation.
+
+The calculation of the diffraction efficiency involves [mode decomposition](Python_User_Interface/#mode-decomposition) of planewaves which is described in [Tutorial/Transmissive Diffraction Spectrum for Planewave at Normal Incidence](Mode_Decomposition.md#transmissive-diffraction-spectrum-for-planewave-at-normal-incidence). An important aspect of setting up this simulation is specifying a diffraction order (planewave) in 2D using the `meep.adjoint.EigenmodeCoefficient` object. This involves three components: (1) the wavevector via the `kpoint_func` parameter, (2) the polarization via the `eig_parity` parameter, and (3) the `eig_vol` parameter must be set to have a *length of one pixel in the periodic direction*. The latter is necessary for MPB to correctly interpret the wavevector in the Cartesian basis rather than the reciprocal-lattice basis defined by the grating period.
+
+The calculation of the derivative of the diffraction efficiency ($F$) with respect to the grating height ($h$) involves using the chain rule to obtain a product of two derivatives (Jacobians): (1) $\partial F / \partial \rho$ and (2) $\partial \rho / \partial h$, where $\rho$ are the density weights (2D grid of $N_x \times N_y = N$ points) used to represent the binary grating as a level set. (1) is the adjoint gradient computed by Meep. (2) requires the level set to be *smoothed* to ensure it that the derivative $\partial \rho / \partial h$ exists. (2) is implemented in Autograd using a custom vector–Jacobian product (vJP) which is used as part of reverse-mode differentiation (backpropagation) to compute $\partial F / \partial h$. An overview of this calculation is shown below.
+
+![](../images/levelset_gradient_backpropagation.png#center)
+
+The derivative $\partial \rho / \partial h$ can be approximated by a finite difference using a one-pixel perturbation applied to the grating height; this is computationally convenient because it greatly simplifies the construction of $\rho(h)$ as explained below. (A finite difference involves two function evaluations of $\rho(h)$, but the cost for this is negligible since it involves no Meep simulations.) The construction of $\rho(h)$ involves two steps: first, constructing a simple binary image $b(h)$ at a high resolution; and second, smoothing $b(h)$ into a continuous level-set function $\rho(h)$.  This smoothing of the image can be performed using a number of different methods including a [signed-distance function](https://en.wikipedia.org/wiki/Signed_distance_function) or convolution filter. In this example, the smoothing is based simply on downsampling the image from a high-resolution grid (10X the resolution of the simulation grid) to the lower-resolution simulation grid using bilinear interpolation, which leads to "gray" pixels at the boundaries between materials in a way that changes continuously with $h$. Only these boundary pixels have nonzero derivatives in the Jacobian $\partial\rho/\partial h$ in this case. This calculation is summarized in the schematic below. 
+
+![](../images/levelset_jacobian_matrix.png#center)
+
+A schematic of the simulation layout showing the design region containing the grating as a levelset is shown below. The adjoint gradient $\partial F / \partial \rho$ computed by Meep at a resolution of 400 pixels/$\mu$m is shown next to this schematic.
+
+![](../images/levelset_adjoint_gradient.png#center)
+
+The simulation script is in [python/examples/adjoint_optimization/binary_grating_levelset.py](https://github.com/NanoComp/meep/tree/master/python/examples/adjoint_optimization/binary_grating_levelset.py). Running this script at five different resolutions produces the output below for the directional derivative computed using the finite difference and the adjoint gradient as well as the relative error from these two results.
+
+```
+RESOLUTION_UM = 50
+dir-deriv:, -0.02535355 (finite difference), -0.01469133 (adjoint), 0.420542 (error)
+
+RESOLUTION_UM = 100
+dir-deriv:, -0.00604817 (finite difference), -0.00473221 (adjoint), 0.217580 (error)
+
+RESOLUTION_UM = 200
+dir-deriv:, -0.00284452 (finite difference), -0.00252470 (adjoint), 0.112432 (error)
+
+RESOLUTION_UM = 400
+dir-deriv:, -0.00140221 (finite difference), -0.00132065 (adjoint), 0.058165 (error)
+
+RESOLUTION_UM = 800
+dir-deriv:, -0.00069606 (finite difference), -0.00067547 (adjoint), 0.029583 (error)
+```
+
+A logarithmic plot of the relative error vs. grid resolution based on these results demonstrates linear convergence. This is expected for fields right on the boundary of a discontinuous interface.
+
+Currently, we recommend using this procedure only for the $\mathcal{S}$ ($E_z$) polarization, since for the for the $\mathcal{P}$ polarization (electric field in the $xy$ plane), there appear to be large discretization errors in the adjoint gradient which [we are currently investigating](https://github.com/NanoComp/meep/pull/2792#issuecomment-2171942477). This is why this demonstration only involved the $\mathcal{S}$ polarization.
+
+![](../images/levelset_adjoint_gradient_error_vs_resolution.png#center)
+
+```py
+from enum import Enum
+from typing import Callable, Tuple
+
+from autograd.extend import primitive, defvjp
+from autograd import numpy as npa
+from autograd import tensor_jacobian_product
+from mpl_toolkits.axes_grid1 import make_axes_locatable
+import matplotlib.pyplot as plt
+import meep as mp
+import meep.adjoint as mpa
+import numpy as np
+from scipy.interpolate import RegularGridInterpolator
+
+
+RESOLUTION_UM = 100
+WAVELENGTH_UM = 0.53
+GRATING_PERIOD_UM = 1.28
+GRATING_HEIGHT_UM = 0.52
+GRATING_DUTY_CYCLE = 0.65
+GRATING_HEIGHT_PERTURBATION_UM = 1.0 / RESOLUTION_UM
+DIFFRACTION_ORDER = 1
+SUBSTRATE_UM = 1.5
+PML_UM = 1.0
+AIR_UM = 1.0
+N_GLASS = 1.5
+DESIGN_REGION_RESOLUTION_UM = 10 * RESOLUTION_UM
+
+Polarization = Enum("Polarization", "S P")
+
+design_region_size = mp.Vector3(
+    GRATING_HEIGHT_UM + GRATING_HEIGHT_PERTURBATION_UM, GRATING_PERIOD_UM, 0
+)
+nx_design_grid = int(design_region_size.x * DESIGN_REGION_RESOLUTION_UM) + 1
+ny_design_grid = int(design_region_size.y * DESIGN_REGION_RESOLUTION_UM) + 1
+nx_sim_grid = int(design_region_size.x * RESOLUTION_UM) + 1
+ny_sim_grid = int(design_region_size.y * RESOLUTION_UM) + 1
+
+DEBUG_OUTPUT = False
+
+
+def design_region_to_meshgrid(nx: int, ny: int) -> Tuple[np.ndarray, np.ndarray]:
+    """Returns the 2D coordinates of the meshgrid for the design region.
+
+    Args:
+      nx, ny: the number of grid points in the x and y directions, respectively.
+
+    Returns:
+      The coordinates of the x and y grid points (2D arrays) as a 2-tuple.
+    """
+
+    xcoord = np.linspace(-0.5 * design_region_size.x, +0.5 * design_region_size.x, nx)
+    ycoord = np.linspace(-0.5 * design_region_size.y, +0.5 * design_region_size.y, ny)
+    xv, yv = np.meshgrid(xcoord, ycoord, indexing="ij")
+
+    return xv, yv
+
+
+@primitive
+def levelset_and_smoothing(grating_height_um: float) -> np.ndarray:
+    """Returns the density weights for a binary grating as a levelset.
+
+    Args:
+      grating_height_um: the height of the grating.
+
+    Returns:
+      The density weights as a flattened (1D) array.
+    """
+
+    xv, yv = design_region_to_meshgrid(nx_design_grid, ny_design_grid)
+    weights = np.where(
+        np.abs(yv) <= 0.5 * GRATING_DUTY_CYCLE * GRATING_PERIOD_UM,
+        np.where(xv <= xv[0][0] + grating_height_um, 1, 0),
+        0,
+    )
+    xcoord = np.linspace(
+        -0.5 * design_region_size.x, +0.5 * design_region_size.x, nx_design_grid
+    )
+    ycoord = np.linspace(
+        -0.5 * design_region_size.y, +0.5 * design_region_size.y, ny_design_grid
+    )
+    interp = RegularGridInterpolator((xcoord, ycoord), weights)
+
+    # Smooth the design weights by downsampling from the design grid
+    # to the simulation grid using bilinear interpolation.
+    sim_grid_xv, sim_grid_yv = design_region_to_meshgrid(nx_sim_grid, ny_sim_grid)
+    smoothed_weights = interp((sim_grid_xv, sim_grid_yv))
+
+    if DEBUG_OUTPUT:
+        fig, ax = plt.subplots()
+        im = ax.imshow(
+            np.transpose(smoothed_weights),
+            cmap="binary",
+            interpolation="none",
+            aspect="equal",
+        )
+        ax.set_title(r"$\rho_{smooth-levelset}(h)$")
+        divider = make_axes_locatable(ax)
+        cax = divider.append_axes("right", size="5%", pad=0.05)
+        fig.colorbar(im, cax=cax)
+        if mp.am_master():
+            fig.savefig("smoothed_levelset.png", dpi=150, bbox_inches="tight")
+
+    return smoothed_weights.flatten()
+
+
+def levelset_and_smoothing_vjp(
+    ans: np.ndarray,
+    grating_height_um: float,
+) -> Callable[[np.ndarray], np.ndarray]:
+    """Returns a function for computing the vector-Jacobian product.
+
+       The Jacobian is computed manually using a finite difference.
+
+    Args:
+      ans: the design weights for the smoothed grating (no perturbation).
+      grating_height_um: the height of the grating.
+
+    Returns:
+      An anonymous function for computing the vector-Jacobian product.
+    """
+
+    xv, yv = design_region_to_meshgrid(nx_design_grid, ny_design_grid)
+    weights = np.where(
+        np.abs(yv) <= 0.5 * GRATING_DUTY_CYCLE * GRATING_PERIOD_UM,
+        np.where(
+            xv <= xv[0][0] + grating_height_um + GRATING_HEIGHT_PERTURBATION_UM, 1, 0
+        ),
+        0,
+    )
+    xcoord = np.linspace(
+        -0.5 * design_region_size.x, +0.5 * design_region_size.x, nx_design_grid
+    )
+    ycoord = np.linspace(
+        -0.5 * design_region_size.y, +0.5 * design_region_size.y, ny_design_grid
+    )
+    interp = RegularGridInterpolator((xcoord, ycoord), weights)
+
+    # Smooth the design weights by downsampling from the design grid
+    # to the simulation grid using bilinear interpolation.
+    xv_sim_grid, yv_sim_grid = design_region_to_meshgrid(nx_sim_grid, ny_sim_grid)
+    smoothed_weights = interp((xv_sim_grid, yv_sim_grid))
+    smoothed_weights = smoothed_weights.flatten()
+
+    jacobian = (smoothed_weights - ans) / GRATING_HEIGHT_PERTURBATION_UM
+
+    if DEBUG_OUTPUT:
+        fig, ax = plt.subplots()
+        im = ax.imshow(
+            np.transpose(jacobian.reshape(nx_sim_grid, ny_sim_grid)),
+            cmap="inferno",
+            interpolation="none",
+            aspect="equal",
+        )
+        ax.set_title(r"$\partial \rho_{smooth-levelset} / \partial h$")
+        divider = make_axes_locatable(ax)
+        cax = divider.append_axes("right", size="5%", pad=0.05)
+        fig.colorbar(im, cax=cax)
+        if mp.am_master():
+            fig.savefig("gradient_smooth_levelset.png", dpi=150, bbox_inches="tight")
+
+    return lambda g: np.tensordot(g, jacobian, axes=1)
+
+
+def grating_1d(pol: Polarization) -> mpa.OptimizationProblem:
+    """Sets up the adjoint optimization of a 1D grating.
+
+    Args:
+      pol: the polarization state (S or P).
+
+    Returns:
+      A meep.adjoint.OptimizationProblem object for the simulation.
+    """
+
+    frequency = 1 / WAVELENGTH_UM
+    pml_layers = [mp.PML(direction=mp.X, thickness=PML_UM)]
+    glass = mp.Medium(index=N_GLASS)
+    size_x_um = (
+        PML_UM
+        + SUBSTRATE_UM
+        + GRATING_HEIGHT_UM
+        + GRATING_HEIGHT_PERTURBATION_UM
+        + AIR_UM
+        + PML_UM
+    )
+    size_y_um = GRATING_PERIOD_UM
+    cell_size = mp.Vector3(size_x_um, size_y_um, 0)
+    k_point = mp.Vector3()
+
+    if pol.name == "S":
+        eig_parity = mp.ODD_Z
+        src_cmpt = mp.Ez
+    else:
+        eig_parity = mp.EVEN_Z
+        src_cmpt = mp.Hz
+
+    src_pt = mp.Vector3(-0.5 * size_x_um + PML_UM, 0, 0)
+    sources = [
+        mp.Source(
+            mp.GaussianSource(frequency, fwidth=0.1 * frequency),
+            component=src_cmpt,
+            center=src_pt,
+            size=mp.Vector3(0, size_y_um, 0),
+        )
+    ]
+
+    matgrid = mp.MaterialGrid(
+        mp.Vector3(nx_sim_grid, ny_sim_grid),
+        mp.air,
+        glass,
+        weights=np.ones((nx_sim_grid, ny_sim_grid)),
+        do_averaging=False,
+    )
+
+    matgrid_region = mpa.DesignRegion(
+        matgrid,
+        volume=mp.Volume(
+            center=mp.Vector3(
+                (-0.5 * size_x_um + PML_UM + SUBSTRATE_UM + 0.5 * design_region_size.x),
+                0,
+                0,
+            ),
+            size=design_region_size,
+        ),
+    )
+
+    geometry = [
+        mp.Block(
+            material=glass,
+            size=mp.Vector3(PML_UM + SUBSTRATE_UM, mp.inf, mp.inf),
+            center=mp.Vector3(-0.5 * size_x_um + 0.5 * (PML_UM + SUBSTRATE_UM), 0, 0),
+        ),
+        mp.Block(
+            material=matgrid,
+            size=matgrid_region.size,
+            center=matgrid_region.center,
+        ),
+    ]
+
+    sim = mp.Simulation(
+        resolution=RESOLUTION_UM,
+        cell_size=cell_size,
+        boundary_layers=pml_layers,
+        k_point=k_point,
+        sources=sources,
+        geometry=geometry,
+    )
+
+    tran_pt = mp.Vector3(0.5 * size_x_um - PML_UM, 0, 0)
+
+    kdiff = mp.Vector3(
+        (frequency**2 - (DIFFRACTION_ORDER / GRATING_PERIOD_UM) ** 2) ** 0.5,
+        DIFFRACTION_ORDER / GRATING_PERIOD_UM,
+        0,
+    )
+
+    obj_args = [
+        mpa.EigenmodeCoefficient(
+            sim,
+            mp.Volume(
+                center=tran_pt,
+                size=mp.Vector3(0, size_y_um, 0),
+            ),
+            mode=1,
+            kpoint_func=lambda *not_used: kdiff,
+            eig_parity=eig_parity,
+            eig_vol=mp.Volume(center=tran_pt, size=mp.Vector3(0, 1 / RESOLUTION_UM, 0)),
+        ),
+    ]
+
+    def obj_func(mode_coeff):
+        return npa.abs(mode_coeff) ** 2
+
+    opt = mpa.OptimizationProblem(
+        simulation=sim,
+        objective_functions=obj_func,
+        objective_arguments=obj_args,
+        design_regions=[matgrid_region],
+        frequencies=[frequency],
+    )
+
+    return opt
+
+
+if __name__ == "__main__":
+    # Only the S polarization is supported (for now).
+    opt = grating_1d(Polarization.S)
+
+    smoothed_design_weights = levelset_and_smoothing(GRATING_HEIGHT_UM)
+
+    obj_val_unperturbed, grad_unperturbed = opt(
+        [smoothed_design_weights], need_gradient=True
+    )
+
+    if DEBUG_OUTPUT:
+        fig, ax = plt.subplots()
+        im = ax.imshow(
+            np.transpose(grad_unperturbed.reshape(nx_sim_grid, ny_sim_grid)),
+            cmap="inferno",
+            interpolation="none",
+            aspect="equal",
+        )
+        ax.set_title(r"$\partial F / \partial \rho_{smooth-levelset}$")
+        divider = make_axes_locatable(ax)
+        cax = divider.append_axes("right", size="5%", pad=0.05)
+        fig.colorbar(im, cax=cax)
+        if mp.am_master():
+            fig.savefig(
+                "gradient_wrt_smoothed_design_weights.png", dpi=150, bbox_inches="tight"
+            )
+
+        fig, ax = plt.subplots()
+        opt.plot2D(init_opt=False, ax=ax)
+        if mp.am_master():
+            fig.savefig("sim_layout.png", dpi=150, bbox_inches="tight")
+
+    defvjp(levelset_and_smoothing, levelset_and_smoothing_vjp)
+    grad_backprop = tensor_jacobian_product(levelset_and_smoothing, 0)(
+        GRATING_HEIGHT_UM,
+        grad_unperturbed,
+    )
+
+    perturbed_design_weights = levelset_and_smoothing(
+        GRATING_HEIGHT_UM + GRATING_HEIGHT_PERTURBATION_UM
+    )
+    perturbed_design_weights = perturbed_design_weights.flatten()
+
+    obj_val_perturbed, _ = opt([perturbed_design_weights], need_gradient=False)
+
+    adj_directional_deriv = GRATING_HEIGHT_PERTURBATION_UM * grad_backprop
+    fnd_directional_deriv = obj_val_perturbed[0] - obj_val_unperturbed[0]
+    rel_err = abs(
+        (fnd_directional_deriv - adj_directional_deriv) / fnd_directional_deriv
+    )
+    print(
+        f"dir-deriv:, {fnd_directional_deriv:.8f} (finite difference), "
+        f"{adj_directional_deriv:.8f} (adjoint), {rel_err:.6f} (error)"
+    )
+```
+
 Compact Notebook Tutorials of Basic Features
 --------------------------------------------
 

diff --git a/doc/docs/images/levelset_adjoint_gradient.png b/doc/docs/images/levelset_adjoint_gradient.png
diff --git a/doc/docs/images/levelset_adjoint_gradient_error_vs_resolution.png b/doc/docs/images/levelset_adjoint_gradient_error_vs_resolution.png
diff --git a/doc/docs/images/levelset_gradient_backpropagation.png b/doc/docs/images/levelset_gradient_backpropagation.png
diff --git a/doc/docs/images/levelset_jacobian_matrix.png b/doc/docs/images/levelset_jacobian_matrix.png