chore(invdes): fixed smoothed projection gradient with beta=inf

[mahlau-flex]

Fixed the gradient computation in smoothed_projection for beta=inf. The problem was that the existing tanh_projection did not handle beta=inf properly.

Greptile Overview

Greptile Summary

Fixed gradient computation in tanh_projection when beta=inf by adding an explicit step function case using np.where, preventing NaN/inf gradients that occurred with the previous tanh formula. The fix enables proper gradient-based optimization in topology design with completely binarized projections.

Key changes:

• Added beta == np.inf special case in tanh_projection (projections.py:70-71) returning step function np.where(array > eta, 1.0, 0.0)
• Refactored tests to use circle_2d_factory pytest fixture eliminating code duplication
• Added comprehensive parametrized test test_projection_gradient_correctness using autograd's check_grads to verify forward and reverse mode gradients up to 2nd order
• Test coverage now includes beta=np.inf case across multiple array sizes, radii, and smoothing parameters

Confidence Score: 4/5

• This PR is safe to merge with minor considerations about edge case documentation
• The fix correctly addresses the gradient computation issue for beta=inf by using a step function instead of the tanh formula that causes numerical instability. Comprehensive test coverage validates gradients across multiple scenarios including the beta=inf case. The only minor concern is documentation clarity around the array == eta edge case behavior.
• No files require special attention - the implementation is sound and well-tested

Important Files Changed

File Analysis

Filename	Score	Overview
tidy3d/plugins/autograd/invdes/projections.py	4/5	Added special case handling for beta=np.inf in tanh_projection to prevent NaN/inf gradients by using step function
tests/test_plugins/autograd/invdes/test_projections.py	5/5	Refactored tests to use fixture for circle generation, added comprehensive gradient correctness tests with parametrized beta=np.inf cases

Sequence Diagram

sequenceDiagram
    participant User
    participant smoothed_projection
    participant tanh_projection
    participant autograd
    
    User->>smoothed_projection: call with array, beta=inf, eta
    smoothed_projection->>tanh_projection: compute original_projected
    
    alt beta == 0
        tanh_projection-->>smoothed_projection: return array (unchanged)
    else beta == inf (NEW FIX)
        tanh_projection->>tanh_projection: np.where(array > eta, 1.0, 0.0)
        tanh_projection-->>smoothed_projection: return step function
    else finite beta
        tanh_projection->>tanh_projection: compute tanh formula
        tanh_projection-->>smoothed_projection: return smooth projection
    end
    
    smoothed_projection->>smoothed_projection: compute gradients & smoothing
    smoothed_projection->>tanh_projection: compute rho_minus_eff_projected
    smoothed_projection->>tanh_projection: compute rho_plus_eff_projected
    smoothed_projection->>smoothed_projection: blend with polynom weights
    smoothed_projection-->>User: return smoothed result
    
    User->>autograd: compute gradient
    autograd->>smoothed_projection: backpropagate
    smoothed_projection->>tanh_projection: gradient through step function (when beta=inf)
    Note over tanh_projection,autograd: np.where has zero gradient<br/>(doesn't cause NaN/inf)
    autograd-->>User: return finite gradients

Loading

[greptile-apps]

_{2 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[yaugenst-flex]

LGTM!

[github-actions]

Diff Coverage

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

• tidy3d/plugins/autograd/invdes/projections.py (100%)

Summary

• Total: 2 lines
• Missing: 0 lines
• Coverage: 100%