Adjoint shape optimization computes the gradient of a figure of merit with respect to the material layout at every point of a design region using only two full-wave Maxwell simulations per iteration, making gradient-based inverse design of photonic components far cheaper than population-based stochastic searches. The method's founding demonstration (Lalau-Keraly, Bhargava, Miller, and Yablonovitch, 2013) on a silicon Y-junction splitter explicitly left three practical issues unresolved, and they remain open: (1) steepest descent is a local optimizer that can be trapped in poor local optima of large, high-dimensional, nonconvex design spaces, and validated strategies for escaping them within realistic simulation budgets are still lacking; (2) multi-stage workflows that optimize in a 2D effective-index approximation before refining in full 3D introduce approximation-induced artifacts that are not quantified or controlled; (3) at insertion losses of order 0.1 dB and below, simulated performance is sensitive to numerical parameters, so cross-study comparisons of ultra-low-loss designs are not yet meaningful. The problem is to resolve these three coupled failure modes of the adjoint design loop — its search reliability, its staging accuracy, and the numerical credibility of its performance claims.

Background

In inverse design of electromagnetic devices, one specifies a figure of merit (FoM), for example the power coupled from an input waveguide into a desired guided mode at an output, and searches a design region for the material layout that maximizes it. The adjoint method computes the derivative of the FoM with respect to the relative permittivity at every point of the design region from just two full-wave Maxwell simulations — one "forward" solve with the physical source and one "adjoint" solve with a source placed at the measurement point or output mode — instead of one solve per parameter as in finite differences; the design geometry is typically moved by the level-set method, in which the boundary between two materials is the zero contour of an auxiliary smooth function, allowing topology changes such as hole nucleation without remeshing. Full-wave solutions are commonly obtained with FDTD (finite-difference time-domain), which discretizes Maxwell's equations on a space-time grid. Because the optimizer is steepest descent on a nonconvex landscape, it converges only to a local optimum. The main competing approach at the time was population-based stochastic search, such as particle swarm optimization (PSO), which evaluates many candidate geometries, each requiring a full-wave simulation.

Lalau-Keraly, Bhargava, Miller, and Yablonovitch (2013) demonstrated this adjoint-level-set workflow on a silicon-photonics Y-junction splitter on a 220-nm-thick silicon-on-insulator platform, optimizing a 2 µm × 2 µm design region with a 200-nm minimum radius of curvature. To limit the cost of full 3D FDTD, they first ran 41 iterations in a 2D effective-index approximation — replacing the 220-nm slab by a planar medium of effective index neff=2.8n_{\mathrm{eff}}=2.8 — and then continued in full 3D, reporting convergence within 51 iterations (102 full-wave simulations) to an insertion loss (output-to-input power ratio, in dB) of 0.07 dB at 1550 nm, remaining below roughly 0.1 dB across 1500-1600 nm; a previously reported manufacturable PSO-optimized splitter with the same material system and feature-size constraints reached a comparable minimum loss but required on the order of 1500 full-wave simulations. When the 2D-optimized geometry was re-simulated in 3D the performance degraded to roughly 0.10-0.12 dB before the 3D refinement recovered it, and no convergence sweeps were reported for the final loss value.

The paper's authors explicitly flagged three unresolved practical issues of this workflow: (i) the adjoint steepest-descent procedure is local, so in large, high-dimensional design spaces it can settle into poor local optima, and they proposed augmenting it with global optimization strategies (for example simulated annealing or other global-search routines) or careful FoM selection; (ii) the 2D effective-index stage is an approximation that does not exactly reproduce full 3D performance, so quantifying and reducing such multi-stage approximation artifacts remained open; and (iii) at very low insertion loss (below about 0.1 dB) simulation results become sensitive to numerical and simulation-parameter choices, so direct comparison of device performance between different optimization studies is unreliable without careful numerical validation. The present problem takes these three issues together as the reliability and verifiability question of adjoint-based electromagnetic shape optimization.

Problem Statement

Resolve the three coupled reliability issues that Lalau-Keraly, Bhargava, Miller, and Yablonovitch (2013) left open for adjoint steepest-descent electromagnetic shape optimization, so that adjoint-designed ultra-low-loss devices can be produced reproducibly and their simulated performance compared across studies with stated confidence:

  1. Local optima in large, high-dimensional design spaces. Quantify the sensitivity of adjoint-gradient optimization to initialization and characterize the distribution and quality of the local optima it can reach for at least two representative electromagnetic device classes (for example waveguide splitters, mode converters, and wavelength demultiplexers on a silicon-on-insulator platform), and develop and validate strategies that demonstrably escape poor local optima within realistic full-wave simulation budgets (at most of the order of 10210^2 full-wave simulations per optimization).

  2. Artifacts of multi-stage reduced-dimensionality optimization. Quantify the approximation-induced artifacts that arise when early iterations optimize in a 2D effective-index representation of a 3D device (in the source demonstration, a 220-nm silicon slab represented by a 2D medium with neff=2.8n_{\mathrm{eff}}=2.8) and the optimization then continues in full 3D, including the observed degradation when the 2D-optimized geometry is re-evaluated in 3D and the possibility that the 2D stage fixes a starting geometry that traps the 3D refinement in an unfavorable basin, and establish criteria or correction methods that control these artifacts for at least two representative device classes.

  3. Numerical validation of ultra-low-loss comparisons. Determine how sensitive computed insertion losses of order 0.1 dB and below are to simulation parameters (mesh resolution, boundary conditions, simulation duration, source and mode-overlap normalization conventions), and establish a validation and uncertainty-quantification protocol under which performance comparisons between different optimization studies and different solvers are meaningful at these loss levels for at least two representative device classes.

References

  1. C. M. Lalau-Keraly, S. Bhargava, O. D. Miller, E. Yablonovitch, "Adjoint shape optimization applied to electromagnetic design", Optics Express 21(18), 21693 (2013). DOI 10.1364/OE.21.021693, https://doi.org/10.1364/oe.21.021693
  2. L. Su, D. Vercruysse, J. Skarda, N. V. Sapra, J. A. Petykiewicz, J. Vučković, "Nanophotonic inverse design with SPINS: Software architecture and practical considerations", Applied Physics Reviews 7, 016101 (2020). DOI 10.1063/1.5131263, arXiv:1910.04829, https://arxiv.org/abs/1910.04829
  3. H. Chung, O. D. Miller, "Tunable Metasurface Inverse Design for 80% Switching Efficiencies and 144° Angular Deflection", ACS Photonics (2020). DOI 10.1021/acsphotonics.0c00787, https://doi.org/10.1021/acsphotonics.0c00787
  4. J. Zhang, S. Bi, G. Zhang, "A directional Gaussian smoothing optimization method for computational inverse design in nanophotonics", arXiv:2007.15079 (2020). https://arxiv.org/abs/2007.15079
  5. V. Nikkhah, A. Pirmoradi, F. Ashtiani, B. Edwards, F. Aflatouni, N. Engheta, "Inverse-designed low-index-contrast structures on a silicon photonics platform for vector–matrix multiplication", Nature Photonics (2024). DOI 10.1038/s41566-024-01394-2, https://doi.org/10.1038/s41566-024-01394-2 (arXiv:2302.00793)
  6. J. R. de Lasson, L. H. Frandsen, P. Gutsche, S. Burger, O. S. Kim, O. Breinbjerg, A. Ivinskaya, F. Wang, O. Sigmund, T. Häyrynen, A. V. Lavrinenko, J. Mørk, N. Gregersen, "Benchmarking five numerical simulation techniques for computing resonance wavelengths and quality factors in photonic crystal membrane line defect cavities", arXiv:1710.02215 (2017); conference version: "Benchmarking five computational methods for analyzing large photonic crystal membrane cavities", NUSOD 2017, DOI 10.1109/NUSOD.2017.8010005, https://arxiv.org/abs/1710.02215
  7. I. A. Krutov, M. Yu. Saygin, I. V. Dyakonov, S. P. Kulik, "Optimized low-loss integrated photonics silicon-nitride Y-branch splitter", AIP Conference Proceedings (2020). DOI 10.1063/5.0011452, https://doi.org/10.1063/5.0011452
  8. P. Dainese, L. Marra, D. Cassara, A. Portes, J. Oh, J. Yang, A. Palmieri, J. R. Rodrigues, A. H. Dorrah, F. Capasso, "Shape optimization for high efficiency metasurfaces: theory and implementation", arXiv:2405.03930 (2024). https://arxiv.org/abs/2405.03930
  9. Z. Liu, J. K. S. Poon, "Comparison of Lumerical FDTD and Tidy3D for three-dimensional FDTD simulations of passive silicon photonic components", arXiv:2506.16665 (2025). https://arxiv.org/abs/2506.16665
  10. M. F. Schubert, "invrs-gym: a toolkit for nanophotonic inverse design research", arXiv:2410.24132 (2024). https://arxiv.org/abs/2410.24132
  11. Y. Zhang, Y. Chen, Y. Chen, Y. Li, X. Zhe, R. Lu, S. Liao, Y. Ma, Z. Guan, "Will Accurate Fields Mislead Photonic Design? From Global Accuracy to Port Readout", arXiv:2606.03038 (2026). https://arxiv.org/abs/2606.03038