Global optimization for inverse design in nanophotonics
P. Bennet1, D. Langevin1, A. Khaireh-Walieh2, O. Teytaud3, P. Wiecha2, A. Moreau1
1UCA, Clermont Auvergne INP, CNRS, Institut Pascal, F-63000, Clermont-Ferrand, France
2LAAS, Université de Toulouse, CNRS, Toulouse, France
3Meta AI Research Paris, France
We present a detailed method for studying the performance of optimization algorithms, particularly in photonics. We define confidence criteria in optimized solutions and emphasize their importance. We show that global optimization algorithms generate solutions meeting these criteria for photonics applications.
Global optimization methods for nanophotonics, despite their high computational cost, are becoming relevant given the performances of modern computers and are yielding promising results. However, in recent years, a large number of optimization algorithms have been proposed, making it difficult to compare their performance on given problems. In this context, we present a practical method to benchmark algorithms in the field of photonics.
Simple observables - convergence curves, consistency curves, and the optimized structure itself - are relevant for judging the reliability of an optimization algorithm and the quality of the solutions generated. Our aim is to define intuitive confidence criteria that can be used to detect whether the result of an optimization is satisfactory. This is particularly important because optimization algorithms can never be sure of finding the true optimal solution, given that they are often non-deterministic or find only local optima.
Here we show how to rely on benchmarks and associated observables to select efficient optimization algorithms, leading to the discovery of novel and satisfactory designs of photonic structures.
Optimization of complex photonic structures with global optimizers