Abstracts

Quantifying the extent of model uncertainty is crucial in the technical feasibility analysis of energy technologies and can provide a significant saving of cost and time. However, performing the uncertainty quantification for a complex chemical process involving coupledPDEs system is computationally prohibitive. Parallel algorithms and parallelization is utilized wherever possible in the entire framework of uncertainty quantification to handlethe involved computational cost. The model complexity, on the other hand, is retained without resorting to any form of reduction or surrogate modeling. The application thatis studied to perform the uncertainty analysis is the post-combustion carbon capture via Rapid Thermal Swing Adsorption using amine sorbents in a hollow fiber contactor. Thermal Swing Adsorption is a dynamic non-isothermal cyclic process with a complex interplayof mass transfer kinetics and equilibrium, and therefore the governing process model is acomplex coupled system of PDEs. The process model developed is initially calibrated using conventional methods of parameter estimation and the performance is benchmarked forcomparison against the results obtained incorporating uncertainties.The computational challenge in performing Bayesian inference, is handled by employing Sequential Monte Carlo, a parallel algorithm based on particle filtering. The uncertainties involved in the process are characterized using four different approaches, viz Hierinf: the data are separated into subsets and the inference is performed for the individual series, Varinflat-inf: the variance of the parametric uncertainties are increased by increasing the variance of the residual errors, Uresvar-inf: wherein, additional model parametersare considered as uncertain in an attempt to reduce the residual variability (errors), Mdiscrep-inf: wherein, the additional uncertainty is introduced in the model structurevia the model discrepancy term. The characterized uncertainties, obtained from each of the four different approaches are propagated through the process model and the uncertainties in the key prediction variables, viz: the product quality and process performance, viz: CO2 swing capacity are obtained. The last component of uncertainty analysis is to be able to design experiments optimally in order to reduce the prediction uncertainties. A newmethod is proposed wherein the prediction uncertainty is reduced through designing experimentsbased on the utility function formulated with the parametric distributions. The proposed method is demonstrated for a simpler system of RTSA, in which only the adsorptionisotherm paramters are considered as uncertain.Advisors/Committee Members: Realff, Matthew J. (committee member), Kawajiri, Yoshiaki (committee member), Jones, Christopher W. (committee member), Lively, Ryan P. (committee member), Vengazhiyil, Roshan J. (committee member).