Machine learning based optimization with applications in optics design, scatterometry, and experiment control

In optics one often targets the optimization of expensive black-box functions. For example, simulation-based design of nano-optic devices involves the time-consuming solution of Maxwell’s equations for different design parameters. Likewise in scatterometry, model-based parameter reconstruction of complex structures requires many expensive numerical simulations of the light scattering process on the structure. Also the optimization of experimental parameters can be regarded as expensive black-box optimization, if the evaluation of the experimental system is time or energy consuming. Bayesian optimization (BO) is a machine-learning (ML) approach that aims to optimize such expensive black-box functions with as few evaluations as possible [1]. To this end, ML models of the objective function are trained, such as Gaussian processes or Bayesian neural networks. By querying the ML models, BO determines parameters that lead to a large expected improvement of the objective value in every iteration. We give an introduction to BO and demonstrate its significant advantages for applications in optics design, scatterometry, and experiment control. [1] ACS Photonics, 6(11), 2726 (2019)