Space-Filling Latin Hypercube Design for Efficient Bayesian Optimization With Application to Semiconductor Development

In this study, we present an initial experimental design for Bayesian optimization (BO) applied to the semiconductor process development. For efficient BO, it can be effective to draw initial sampling nodes that cover the search space as uniformly as possible. In particular, in process optimization, it is necessary to appropriately place the sampling nodes within the high-dimensional search space consisting of multiple process parameters. To address this issue, we propose a space-filling Latin hypercube design (LHD) with a small fill distance and a large separation radius based on simulated annealing. Compared with Maximin LHD and Sobol’ design, the proposed design exhibits the smallest fill distance and a large separation radius, which is equivalent to that of Maximin LHD. Through an example focused on optimizing high aspect ratio hole etching, we demonstrate that the proposed design effectively reduces the number of experiments required for the convergence of BO compared to conventional designs. Additionally, it can be applied to various processes such as chemical vapor deposition, atomic layer deposition and wet etching, among others.