Since complex simulation contains many factors and responses that interact in a nonlinear manner, it is important to use metamodeling for representing the causal relationships within the simulation models in a compact way. With the selected mathematical structure, the effectiveness of metamodeling depends on the comprehensiveness of the training data that is closely related to the size of the scenario space and available computing resources. Generally, sequential experimental design methods are more efficient than one-shot ones, but the later depend on the domain knowledge of the experimenters and are uneasy to be conducted automatically. This paper proposes a sequential neighbor exploratory experimental design (SNEED) method for metamodeling purpose. Through the peaks function example, we compare this new method to Latin hypercube with a support vector regression metamodel trained by their training data respectively. The result shows that under the same experiment sample count, the SNEED method produces better regression performance.