Identifiability-based model decomposition for hierarchical calibration

Abstract

The computational model has become an essential tool in many engineering applications. To take full advantage of a computational model, its accuracy must be guaranteed. Validation and verification (V&V) methods have been proposed to enable construction of accurate computational models. One of the challenges in V&V is to calibrate the large number of parameters of the model of interest. To effectively calibrate the large number of calibration parameters, a hierarchical calibration method has been proposed; this method decomposes the system model into smaller models and calibrates the parameters for the decomposed models. However, the method uses only qualitative criteria for model decomposition in the hierarchical calibration. This study identifies the problems of using only these qualitative criteria for hierarchical calibration, and instead proposes to use identifiability, obtained from the Fisher information matrix, as a quantitative criterion. A detailed process for the proposed identifiability-based model decomposition is described. During decomposition, the experiments are designed to meet identifiability requirements. The proposed decomposition method is verified by examining the Pratt truss bridge model as a case study; the method shows accurate calibration results and requires fewer experiments, as compared with two arbitrary decompositions.