Multi-scale modeling of an amine sorbent fluidized bed adsorber with dynamic discrepancy reduced modeling

Abstract

Inaccuracy in chemical kinetic models is a problem that affects the reliability of large-scale process models. Detailed and accurate kinetic models could improve this situation, but such models are often too computationally intensive to utilize at process scales. This work presents a method for bridging the kinetic and process scales while incorporating experimental data in a Bayesian framework. “Dynamic Discrepancy Reduced Modeling” (DDRM) enables the use of complex chemical kinetic models at the process scale. DDRM inserts Gaussian process (GP) stochastic functions into the first-principles dynamic model form, enabling a sharp reduction in model order. Uncertainty introduced through the order reduction is quantified through Bayesian calibration to bench-scale experimental data. The use of stochastic functions within the dynamic chemical kinetic model enables this uncertainty to be correctly propagated to the process scale. The model reduction and uncertainty quantification framework was applied to a reaction–diffusion model of mesoporous silica-supported, amine impregnated sorbents; model order reduction of over an order of magnitude was achieved through a sharp reduction in the resolution of the approximate solution for diffusion. GPs of the Bayesian smoothing splines analysis of variance (BSS-ANOVA) variety were applied to the diffusion coefficients and equilibrium constants in the model. The reduced model was calibrated using experimental dynamic thermogravimetric data, with prior probability distributions for physical model parameters derived from quantum chemical calculations. Model and parameter uncertainties represented in the posterior distribution were propagated to a bubbling fluidized-bed adsorber simulation implemented in Aspen Custom Modeler. This work is the first significant demonstration of the dynamic discrepancy technique to produce robust reduced order models for bench-to-process multi-scale modeling and serves as a proof-of-concept.