Illustrated tutorial on global optimization in nanophotonics

Pauline Bennet, Denis Langevin, Chaymae Essoual, Abdourahman Khaireh-Walieh, Olivier Teytaud, Peter R. Wiecha, Moreau, AntoineJournal of the Optical Society of America B 2023 · arXiv

This tutorial paper serves as a practical guide for researchers in nanophotonics looking to navigate the often-murky waters of global optimization. By combining rigorous benchmarking with physical intuition, the authors provide a roadmap for "inverse design"—the process of starting with a desired optical response and letting an algorithm find the structure that produces it.

TL;DR

The paper provides a methodology and open-source toolkit (PyMoosh and Nevergrad) to solve the "reproducibility crisis" in photonic optimization. It identifies Differential Evolution (DE) as a robust "workhorse" algorithm and demonstrates that physical regularity in a solution is often a hallmark of a successful global search.

What Problem Does This Paper Address?

Designing nanophotonic structures (like solar cell coatings or optical filters) is notoriously difficult because light behaves like a wave. This wave nature creates a "cost function landscape" littered with local minima—trap-like solutions that look good but aren't the best possible design.

While local optimization (like gradient descent) is fast, it almost always gets stuck in these traps. Global optimization can find the true "best" design, but the field suffers from a lack of standardized benchmarking and reproducibility. Physicists often treat these algorithms as "black boxes," leading to results that are hard to verify or apply to different problems.

Methods

The authors advocate for Parametric Optimization (PO), where the number of design variables is kept relatively low (typically under 100). They utilize two primary software tools:

  • PyMoosh: A Python library for simulating the optical properties of multilayered structures.
  • Nevergrad: An open-source benchmarking library from Meta AI that provides a unified interface for dozens of optimization algorithms.

The paper highlights Differential Evolution (DE) and its variants, such as Quasi-Oppositional DE (QODE). Unlike gradient-based methods, these "population-based" algorithms maintain a pool of candidate designs that "evolve" over time, allowing the search to jump out of local traps.

Datasets (Test Cases)

The authors introduce five distinct photonic benchmarks, provided in an online repository for the community:

  1. Bragg Mirror: Maximizing reflectivity for a multilayer stack.
  2. Ellipsometry: Retrieving the thickness and refractive index of a thin film from experimental data.
  3. Photovoltaics: Designing an antireflective coating to maximize solar cell current.
  4. 2D Grating: A complex structure inspired by Morpho butterfly wings to control light diffraction.
  5. 3D Nanoantenna: Optimizing the positions of gold nanocubes to steer light emission.

Key Insights and Ideas

  • The "Aha" Moment of Regularity: The authors observe that when a global optimization truly succeeds, the resulting design often "spontaneously" exhibits physical regularity or symmetry (e.g., periodic layers). If an optimized design looks messy or random, it is often a sign that the algorithm got stuck in a local minimum.
  • Consistency Curves: The paper introduces "consistency curves" as a vital tool. By running an optimization multiple times and plotting the results sorted by quality, a "plateau" at the lowest value gives the researcher statistical confidence that they have found the true global optimum.
  • Initialization Matters: A simple mathematical trick—initializing the population with "quasi-oppositional" points (adding a symmetric counterpart to every random point)—significantly boosts the efficiency of the search.
  • DE is "Modular": Differential Evolution is particularly suited for photonics because it tends to preserve "chunks" of parameters that work well together, which physically corresponds to functional blocks of a photonic structure.

Key Results

  • DE vs. The Rest: In head-to-head benchmarks, Differential Evolution variants (QODE and QNDE) consistently outperformed the popular CMA-ES (Covariance Matrix Adaptation Evolution Strategy) across most photonic test cases.
  • Gradient Failure: Local gradient-based methods (like BFGS) were shown to be completely unsuited for complex problems like the 3D nanoantenna, where the "landscape" is too rugged.
  • Reproducibility: By providing Jupyter notebooks, the authors demonstrate that complex inverse design can be made accessible and verifiable for any researcher with basic Python knowledge.

Limitations and Future Work

  • The Curse of Dimensionality: While the methods work excellently for dozens of parameters, they struggle to scale to the thousands of parameters used in "Topology Optimization" (TO). Bridging the gap between global BBO and high-dimensional TO remains a challenge.
  • Computational Expense: Global optimization is still computationally "expensive." The authors suggest that future work should focus on leveraging parallel computing (MPI) more effectively to handle 3D simulations.
  • Noise Handling: Real-world factors like manufacturing tolerances and noisy solar spectra can confuse algorithms; the authors note that more work is needed to integrate "robustness" directly into the optimization process.

Summary updated

Abstract

Numerical optimization for the inverse design of photonic structures is a tool which is providing increasingly convincing results -- even though the wave nature of problems in photonics makes them particularly complex. In the meantime, the field of global optimization is rapidly evolving but is prone to reproducibility problems, making it harder to identify the right algorithms to use. This paper is thought as a tutorial on global optimization for photonic problems. We provide a general background on global optimization algorithms and a rigorous methodology for a physicist interested in using these tools -- especially in the context of inverse design. We suggest algorithms and provide explanations for their efficiency. We provide codes and examples as an illustration than can be run online, integrating quick simulation code and Nevergrad, a state-of-the-art benchmarking library. Finally, we show how physical intuition can be used to discuss optimization results and to determine whether the solutions are satisfactory or not.
BibTeX
@article{bennet2023illustrated,
  author = {Pauline Bennet and Denis Langevin and Chaymae Essoual and Abdourahman Khaireh-Walieh and Olivier Teytaud and Peter R. Wiecha and Moreau, Antoine},
  title = {{Illustrated tutorial on global optimization in nanophotonics}},
  journal = {Journal of the Optical Society of America B},
  year = {2023},
}
Figure