Bayesian Target‐Vector Optimization for Efficient Parameter Reconstruction

Parameter reconstructions are indispensable in metrology. Here, the objective is to explain K experimental measurements by fitting them a parameterized model of measurement process. The parameters regularly determined least-square methods, that is, minimizing sum squared residuals between predictions and observations, χ2. functions often involve computationally demanding numerical simulations. Bayesian optimization methods specifically suited for expensive functions. However, contrast such as Levenberg–Marquardt algorithm, they only take value χ2 into account, neglect individual outputs. A target-vector scheme with improved performance over previous developments, considers all contributions function parameter reconstruction problems which based on hundreds observations presented. Its compared established an optical metrology problem two synthetic least-squares problems. proposed method outperforms methods. It also enables determine accurate uncertainty estimates very few actual using Markov chain Monte Carlo sampling trained surrogate model.