(An) efficient method for the selection of local surrogates

Abstract

In engineering, physical systems are typically described by known functions, but such functions aren’t always available in all contexts. Unknown functions that link inputs to outputs are referred to as ”black box functions”, and are obtained from expensive experiments or simulations. The development of surrogate models becomes relevant in scenarios where balancing accuracy and efficiency is crucial. These metamodels enable rapid analysis at the tradeoff of a slight, but hopefully acceptable, reduction in accuracy. This thesis aims to develop an efficient method for selecting multiple local surrogate models. It addresses the challenges and limitations inherent in existing methodologies and proposes a novel approach to enhance accuracy and efficiency in surrogate model selection. The research employs a combination of artificial neural networks and Kriging techniques. The methodology and algorithms for selecting and optimising local surrogate models are introduced, considering factors like computational cost, accuracy, and model complexity. The proposed methodology is tested through the engineering application focusing on the design methodology of urban transit vehicles. The developed method offers a more efficient and accurate approach to selecting local surrogate models, with potential implications for a wide range of engineering applications.