Mukhtar Azfarizal, Yasir Ahmad Shah Hizam Md, Nasir Mohamad Fariz Mohamed
Institute of Sustainable Energy, Putrajaya Campus, Universiti Tenaga Nasional, Jalan IKRAM-UNITEN, 43000 Kajang, Malaysia.
College of Engineering, Putrajaya Campus, Universiti Tenaga Nasional, Jalan IKRAM-UNITEN, 43000 Kajang, Malaysia.
Heliyon. 2023 Jul 26;9(8):e18674. doi: 10.1016/j.heliyon.2023.e18674. eCollection 2023 Aug.
Complex computer codes are frequently used in engineering to generate outputs based on inputs, which can make it difficult for designers to understand the relationship between inputs and outputs and to determine the best input values. One solution to this issue is to use design of experiments (DOE) in combination with surrogate models. However, there is a lack of guidance on how to select the appropriate model for a given data set. This study compares two surrogate modelling techniques, polynomial regression (PR) and kriging-based models, and analyses critical issues in design optimisation, such as DOE selection, design sensitivity, and model adequacy. The study concludes that PR is more efficient for model generation, while kriging-based models are better for assessing max-min search results due to their ability to predict a broader range of objective values. The number and location of design points can affect the performance of the model, and the error of kriging-based models is lower than that of PR. Furthermore, design sensitivity information is important for improving surrogate model efficiency, and PR is better suited to determining the design variable with the greatest impact on response. The findings of this study will be valuable to engineering simulation practitioners and researchers by providing insight into the selection of appropriate surrogate models. All in all, the study demonstrates surrogate modelling techniques can be used to solve complex engineering problems effectively.
复杂的计算机代码在工程中经常被用来根据输入生成输出，这可能使设计师难以理解输入和输出之间的关系，也难以确定最佳输入值。解决这个问题的一种方法是将实验设计（DOE）与代理模型结合使用。然而，对于如何为给定数据集选择合适的模型，缺乏相关指导。本研究比较了两种代理建模技术，即多项式回归（PR）和基于克里金法的模型，并分析了设计优化中的关键问题，如DOE选择、设计敏感性和模型充分性。研究得出结论，PR在模型生成方面更有效，而基于克里金法的模型由于能够预测更广泛的目标值范围，在评估最大-最小搜索结果方面表现更好。设计点的数量和位置会影响模型的性能，基于克里金法的模型的误差低于PR。此外，设计敏感性信息对于提高代理模型效率很重要，PR更适合确定对响应影响最大的设计变量。本研究的结果将为工程模拟从业者和研究人员提供有价值的见解，帮助他们选择合适的代理模型。总而言之，该研究表明代理建模技术可用于有效解决复杂的工程问题。
