Computational model validation under uncertainty

This paper develops a methodology to assess the validity of computation models under uncertainty using the concept of Bayesian hypothesis testing, by comparing model prediction and experimental observation. The concept of model reliability is introduced by comparing the prior and posterior distributions of the error. The proposed method is illustrated by considering discretization error in the case of finite element-based computational models. A collocation-based stochastic response surface method is developed for efficiency in computing the stochastic response and error. The paper extends the above approach to assess the validity of large-scale computational models by combining system reliability concepts with a Bayesian model validation approach. While full system testing may be impossible, component-level testing may be possible to validate smaller modules of the computational model. Bayes networks are used to propagate he validation information from the component modules to system-level prediction.