A Super-Metamodeling Framework to Optimize System Predictability

Abstract

Statistical metamodels can robustly predict manufacturing process and engineering systems design results. Various techniques, such as Kriging, polynomial regression, artificial neural network and others, are each best suited for different scenarios that can range across a design space. Thus, methods are needed to identify the most appropriate metamodel or model composite for a given problem. To account for pros and cons of different metamodeling techniques for a wide diversity of data sets, in this paper we introduce a super-metamodel optimization framework (SMOF) to improve overall prediction accuracy by integrating different metamodeling techniques without a need for additional data. The SMOF defines an iterative process first to construct multiple metamodels using different methods and then aggregate them into a weighted composite and finally optimize the super-metamodel through advanced sampling. The optimized super-metamodel can reduce an overall prediction error and sustains the performance regardless of dataset variation. To verify the method, we apply it to 24 test problems representing various scenarios. A case study conducted with additive manufacturing process data shows method effectiveness in practice.