Global Optimization Using Mixed Surrogates and Space Elimination in Computationally Intensive Engineering Designs

Surrogate-based modeling is an effective search method for global design optimization over well-defined areas using complex and computationally intensive analysis and simulation tools. However, indentifying the appreciate surrogate models and their suitable areas remains a challenge that requires extensive human intervention. In this work, a new global optimization algorithm, namely Mixed Surrogate and Space Elimination (MSSE) method, is introduced. Representative surrogate models, including Quadratic Response Surface, Radial Basis function, and Kriging, are mixed with different weight ratios to form an adaptive metamodel with best tested performance. The approach divides the field of interest into several unimodal regions; identifies and ranks the regions that likely contain the global minimum; fits the weighted surrogate models over each promising region using additional design experiment data points from Latin Hypercube Designs and adjusts the weights according to the performance of each model; identifies its minimum and removes the processed region; and moves to the next most promising region until all regions are processed and the global optimum is identified. The proposed algorithm was tested using several benchmark problems for global optimization and compared with several widely used space exploration global optimization algorithms, showing reduced computation efforts, robust performance and comparable search accuracy, making the proposed method an excellent tool for computationally intensive global design optimization problems.