A model refinement framework for statistical model validation

Abstract

As the importance of virtual testing has been increased for cost-effective product design and design evaluation, researchers focus on studying validation and verification (V&V) to increase the computational model predictability. Model validation process can make the computational model accurately through the model calibration and validity check process

however, in some cases, unacknowledged uncertainties such as lack of knowledge and human mistakes still exist and decrease the predictability of the model. To overcome this challenge, this thesis presents a model refinement framework for statistical model validation. This framework consists of the three steps

1) invalidity analysis, 2) invalidity reasoning tree (IRT) and 3) invalidity sensitivity study. Invalidity analysis seeks possible causes for invalidity. Then, the IRT determines a parametric form of refinement candidates from the possible causes and invalidity sensitivity analysis finally checks the effect of the candidates quantitatively. Model calibration and validity check are followed to ensure good model predictability. The proposed method is demonstrated with the TFT-LCD fracture of a smartphone.