feat(invdes): smoothed projection

[mahlau-flex]

Smoothed projection from the paper https://arxiv.org/pdf/2503.20189

Greptile Overview

Greptile Summary

This PR adds smoothed_projection, a new inverse design projection function that supports subpixel smoothing with β→∞, based on the paper by Alec Hammond (arxiv.org/pdf/2503.20189). The implementation integrates out discontinuities in the projection function, allowing smooth gradient flow even at infinite projection strength.

Key changes:

• Added smoothed_projection function in projections.py with comprehensive documentation
• Updated API documentation to include the new function
• Added test coverage for β=∞ case
• Properly exported the function in module __all__

Issues found:

• Minor typos in docstrings and comments that need correction
• Missing CHANGELOG entry for this user-facing feature addition

Confidence Score: 4/5

• This PR is safe to merge with minor corrections needed for typos and CHANGELOG entry
• The implementation is mathematically sound, well-documented, and includes basic test coverage. The function follows established patterns from other projection functions in the codebase. Only minor issues found: spelling errors in docstrings and missing CHANGELOG entry per project guidelines.
• All files are acceptable but tidy3d/plugins/autograd/invdes/projections.py needs typo corrections

Important Files Changed

File Analysis

Filename	Score	Overview
tidy3d/plugins/autograd/invdes/projections.py	4/5	Added smoothed_projection function for subpixel smoothing with β→∞ support, well-documented with minor typo
tests/test_plugins/autograd/invdes/test_projections.py	4/5	Added basic test for smoothed_projection with beta=inf, validates discrete output behavior
tidy3d/plugins/autograd/invdes/init.py	5/5	Added smoothed_projection to exports and all list
docs/api/plugins/autograd.rst	5/5	Added documentation reference for smoothed_projection

Sequence Diagram

sequenceDiagram
    participant User
    participant smoothed_projection
    participant tanh_projection
    participant np_gradient
    
    User->>smoothed_projection: array, beta, eta
    Note over smoothed_projection: Assert array.ndim == 2
    smoothed_projection->>tanh_projection: array, beta, eta
    tanh_projection-->>smoothed_projection: rho_projected
    
    smoothed_projection->>np_gradient: array
    np_gradient-->>smoothed_projection: rho_filtered_grad
    
    Note over smoothed_projection: Compute gradient norm<br/>and sanitize for divide-by-zero
    Note over smoothed_projection: Calculate distance d to interface<br/>d = (eta - array) / grad_norm
    Note over smoothed_projection: Determine needs_smoothing:<br/>interface within voxel?
    
    Note over smoothed_projection: Compute fill factors F and F_minus<br/>using analytic expressions
    Note over smoothed_projection: Calculate bounds:<br/>rho_filtered_minus, rho_filtered_plus
    
    smoothed_projection->>tanh_projection: rho_filtered_minus, beta, eta
    tanh_projection-->>smoothed_projection: rho_minus_eff_projected
    smoothed_projection->>tanh_projection: rho_filtered_plus, beta, eta
    tanh_projection-->>smoothed_projection: rho_plus_eff_projected
    
    Note over smoothed_projection: Apply smoothing at interfaces<br/>rho_projected_smoothed = (1-F)*minus + F*plus
    Note over smoothed_projection: Use np.where to select:<br/>smoothed or standard projection
    
    smoothed_projection-->>User: smoothed result

Loading

Context used:

• Rule from dashboard - Require a changelog entry for any PR that is not purely an internal refactor. (source)

[greptile-apps]

_{4 files reviewed, 3 comments}

_{Edit Code Review Agent Settings | Greptile}

[github-actions]

Diff Coverage

Diff: origin/develop...HEAD, staged and unstaged changes

• tidy3d/plugins/autograd/invdes/init.py (100%)
• tidy3d/plugins/autograd/invdes/projections.py (95.7%): Missing lines 137

Summary

• Total: 24 lines
• Missing: 1 line
• Coverage: 95%

tidy3d/plugins/autograd/invdes/projections.py

Lines 133-141

  133        (2025).
  134     """
  135     # sanity checks
  136     if array.ndim != 2:
! 137         raise ValueError(f"Smoothed projection expects a 2d-array, but got shape {array.shape=}")
  138 
  139     # smoothing kernel is circle (or ellipse for non-uniform grid)
  140     # we choose smoothing kernel with unit area, which is r~=0.56, a bit larger than (arbitrary) default r=0.55 in paper
  141     dx = dy = scaling_factor

---

[yaugenst-flex]

Thanks @mahlau-flex this is really nice to have!

[yaugenst-flex]

Thanks @mahlau-flex looks great! Only minor formatting nits on the docstring.

[yaugenst-flex]

LGTM!