LAUSR

Abstract In the design of molecules, materials, and processes, adaptive design of experiments (ADoE) is conducted to minimize the number of experiments. Although Bayesian optimization (BO) is an effective tool, BO merely selects a candidate from a limited number of samples, and the samples do not necessarily contain the optimal solution. Furthermore, because upper and lower limits are set for explanatory variables X, it is not possible to obtain solutions that go beyond these limits. To solve these issues, an approach to ADoE called Gaussian mixture regression-based optimization (GMRBO) is proposed. Because GMR models can estimate the X values directly based on the target value of the objective variable y, the optimal solution for X can be calculated without having to establish upper and lower limits to X. GMRBO can allow the target y value to be achieved with a dramatically smaller number of experiments than by BO, especially when the number of X-variables is large.