Scale Bridging and Uncertainty Propagation in Chemical Process Modeling with Bayesian Nonparametric Regression

Bayesian nonparametric regression is a powerful method for building reduced models of chemical reactions and distributed systems.   Motivated by Takens' theorem in dynamic systems topology, Gaussian process-based stochastic functions are insinuated into chemical system models, leading to a stochastic dynamic system of drastically reduced order.  Karhunen-Loeve decomposition of the GP function kernels leads to calibration of the models to training data sets using standard approaches.  Low-order, calibrated stochastic models then serve as vehicles for propagation of uncertainty across modeling length scales.  This methodology can be applied to problems in design of bench and pilot-scale experiments for process design, machine learning in chemical process control, or in other settings where reduced modeling is required.  Benchmarking examples derived from amine-based carbon capture and steam reformation of methane will be presented.