LAUSR

As parallel computing becomes increasingly important in many real-world applications, a batch sequential experimental design (BSED), which adds a batch of computer experiments per iteration and runs these simulations in parallel, is gaining popularity in surrogate modeling. This article proposes sequential online dispatch in design of experiments (SODDE) for single- and multiple-response surrogate modeling when multiple processors work in parallel but not independently. The proposed method includes several unique features: 1) it works with any popular acquisition function to select a single new sample point at each iteration; 2) it minimizes the idle time of all processors; 3) it rapidly updates the surrogate model; and 4) it dynamically reconstructs the surrogate model when a simulation process aborts, minimizing the impact incurred by the abortion. The effectiveness of SODDE is evaluated in one mathematical example and one industrial problem. The latter problem considers blade design of horizontal axis wind turbines (HAWTs). The cost of finding blade geometry that results in the desired aerodynamic behavior of HAWTs is estimated for BSED and SODDE. Relative to BSED, SODDE reduces the costs by up to approximately 31%. Note to Practitioners—Computer simulation models are often used to represent the behavior of complex engineered systems, such as automobiles, aircraft, and wind turbines. Although running computer simulations allows analyzing the performance of these systems, the simulations can still be computationally costly (e.g., hours of runtime for numerically solving a large number of differential equations in one simulation), which makes the performance analysis task challenging. It is important to design an innovative simulation process that replaces the original expensive simulation models by cheap-to-build surrogate models, and at the same time, minimize the computational costs for building the surrogate models. The new method developed in this article allows experimental designers to run batch sequential iterations and run these iterations in parallel. This, in engineering practice, will help meet the need for long and complex experimental designs scenarios in a parallel computing environment, where the computer “up time” (or runtime) needs to be maximized to save time and ultimately, money. Cost analysis results from an industrial problem involving the inverse design of wind turbine blades suggest that the proposed method reduces the computational costs by up to approximately 31% over an existing method.