Adaptive hybrid surrogate modeling for complex systems

This paper explores the effectiveness of the recently developed surrogate modeling method, the Adaptive Hybrid Functions, when applied in conjunction with different sampling techniques and different sample sizes. The Adaptive Hybrid Functions is a hybrid surrogate modeling method that seeks to exploit the advantages of each component surrogate. To explore its effectiveness, the Adaptive Hybrid Functions is applied to model four disparate complex engineering systems, namely: 1) wind farm power generation, 2) product platform planning (for universal electric motors), 3) three-pane window heat transfer, and 4) onshore wind farm cost estimation. The effectiveness of the following three distinct sampling techniques is investigated: 1) Latin hypercube sampling, 2) Sobol's quasi-random sequence, and 3) the Hammersley sequence sampling. Cross-validation is used to evaluate the accuracy of the resulting surrogate models. It was observed that the Sobol's sequence and the Latin hypercube sampling techniques provided better accuracy in the case of high-dimensional problems, whereas the Hammersley sequence sampling technique performed better in the case of low-dimensional problems. It was further observed that a monotonic increase in modeling accuracy is not necessarily accomplished by simply increasing the sample size for different problems. Overall, the Adaptive Hybrid Functions provided acceptable to high accuracy in representing complex system behavior.