The complexities involved in an optimization problem at a system level require knowledge base that has information on different approaches and customization of these approaches to a specific class of the optimization problems. One approach that is commonly used is the metamodel based design optimization. The metamodel is 1) a conceptual model for capturing, in abstract terms, essential characteristics of a given optimization problem, and 2) a schema of sufficient formality to enable the problem modeled to be serialized to statements in a concrete optimization language [1]. Optimization is performed based on this metamodel. This metamodel approach has been proven effective and accurate in providing the global optimum. Depending upon the computational hardware availability in an organization, the metamodel based optimization could be much faster way of achieving the optimized solution.
However, the accuracy of the optimization is highly dependent on the quality of metamodel generated. The quality of a metamodel has strong dependency on the space-filling properties of the Design of Experiments (DoE). The accuracy of metamodel can be significantly improved by better space filling or discretization of the design space.
This paper contrasts different approaches of evolving metamodels and compares their ability to improve the accuracy of the optimized solution. The paper demonstrates assessment of 1) the Group-based space-filling algorithm that creates DoEs that have groups of points with space-filling properties that allows the engineer to create a metamodel sequentially and adding points only when it is needed, 2) Custom development in commercial code modeFRONTIER, 3) Adaptive sampling in commercial code iSight. The load case considered for this study is the 35mph, flat frontal event with design variables being gages on structural load paths. The responses tracked included vehicle pulse and intrusions. The paper contrasts the benefits provided by these approaches and types of evolution needed while working with them.