Adaptive Codomain Sampling for Accurate Surrogate Modeling

Data-driven surrogate models are extensively used to approximate expensive black-box target functions, aiming to accelerate complex engineering design processes. The fidelity of surrogate models hinges on well-distributed training dataset. Most existing research can generate training samples that are uniformly distributed over the domain of the target function. However, the sample distribution in the codomain of the target function is more complex and unknown. Creating a dataset that meets the distribution requirements in the codomain is a significant challenge.In this paper, we propose an adaptive sampling method to generate datasets for training surrogate models. Compared to the conventional sampling method, the proposed method can generate samples that provide a uniform and broad coverage over the codomain, while maintaining relatively uniform coverage of the domain. Consequently, the dataset generated by the proposed method significantly enhances the accuracy of surrogate models. The effectiveness of the proposed method is validated by its application in optimizing V-shaped permanent magnet synchronous machines.