Bayesian optimization and rigorous modelling of a highly efficient 3D metamaterial mode converter

Abstract

We combine a statistical learning-based global optimization strategy with a high order 3D Discontinuous Galerkin Time-Domain (DGTD) solver to design a compact and highly efficient graded index photonic metalens. The metalens is composed of silicon (Si) strips of varying widths (in the transverse direction) and lengths (in the propagation direction) and operates at the telecommunication wavelength. In our work, we tackle the challenging Transverse Electric case (TE) where the incident electric field is polarized perpendicular to strips direction. We reveal that the focusing efficiency approaches 80% for the traditional design with fixed strip lengths and varying widths. Nevertheless, we demonstrate numerically that the efficiency is as high as 87% for a design with varying strip lengths along the propagation direction.

Strength through defects: A novel Bayesian approach for the optimization of architected materials

“…Bayesian optimization (BO) has proven to work well for optimization problems with costly black-box functions by finding global optimum with minimum number of function calls (32,33). However, in previous studies, BO has been used for architected materials for problems with continuous variables (30,34,35). We introduce a novel BO framework, which we term "evolutionary Monte Carlo sampling" (EMCS), to optimize the prohibitively expensive qualitative input design space of our architected microlattice structures.…”

Multiobjective Statistical Learning Optimization of RGB Metalens

“…DGTD can be viewed as a mixture of a classical (continuous) finite element time domain (FETD) and a finite volume time domain (FVTD), which turns out to be well adapted for the simulation of nanoscale light–matter interaction problems. , The choice of time domain solver is fully justified for multiwavelength metasurface designs, as all objective values (device performance in a wavelength range) are obtained with a single simulation run. As it accounts for the full metasurface structure or elementary supercell unit, the solver is assessing near-field coupling between the neighboring elements, which is a fundamental factor in designing highly efficient metasurfaces. , Finally, due to the adaptability of our high-order DGTD solver in handling large-scale problems, the number of mesh cells is considerably reduced in comparison to other time-domain methods such as the finite difference time domain (FDTD).…”

“…As it accounts for the full metasurface structure or elementary supercell unit, the solver is assessing near-field coupling between the neighboring elements, which is a fundamental factor in designing highly efficient metasurfaces. 16,78 Finally, due to the adaptability of our high-order DGTD solver in handling large-scale problems, the number of mesh cells is considerably reduced in comparison to other time-domain methods such as the finite difference time domain (FDTD).…”