Sequential Bayesian adaptive Monte Carlo model discrimination framework with application to chemical kinetics. (1st December 2015)

• Record Type: Journal Article
• Authors: Masoumi, Samira; Reilly, Park M.; Duever, Thomas A.
• Is Part Of: Chemical engineering science. Volume 137(2015)
• Page Start: 796
• Page End: 806
• Publication Date: 2015-12-01
• Subjects: Model selection -- Sequential Bayesian -- Design of experiment -- Model uncertainty -- Adaptive MCMC -- Model discrimination; Chemical engineering -- Periodicals; Génie chimique -- Périodiques; Chemical engineering; Periodicals; Electronic journals; 660
• DOI: 10.1016/j.ces.2015.06.067
• Languages: English
• ISSNs: 0009-2509

Abstract

Abstract: Sequential Bayesian Monte Carlo Model Discrimination (SBMCMD) framework has been previously proposed by the authors for the purpose of determining the underlying mechanisms of a system such as a chemical reaction (Masoumi et al., 2013 ). SBMCMD relies on sampling from the model parameters distribution using Markov Chain Monte Carlo, MCMC, methods. Effective tuning of MCMC methods, when applying to some nonlinear models, can be tedious and challenging. This limits using SBMCMD in many practical applications. The aim of this paper is to address this limitation and facilitate exploiting of the proposed framework with regards to nonlinear structured models. This is achieved by using adaptive random-walk Metropolis–Hasting method for sampling from the models parameter. This method is an adaptive MCMC algorithm that takes care of adjusting its parameters automatically. Two implementations of the adaptive SBMCMD framework have been presented and applied to case studies. Results of two implementations have been compared, and the effect of preliminary data has been discussed. Highlights: SBMCMD framework is a combination of model selection and design of experiment (DOE). SBMCMD has the ability of handling nonlinear systems using MCMC methods. Adaptive MCMC sampling method has been used in the new implementations of SBMCMD. Applying DOE lets to discriminates models with the minimum number of experiments. The effect of the preliminary data and experimental error has been discussed.