Bayesian microkinetic modeling of epoxy resin curing and water gas shift catalysis.

摘要

This work focuses on describing a framework for building forward models as part of a system for computer aided materials discovery called "discovery informatics." The methodology is a data intensive search for Quantitative Property Activity Relationships (QPAR) capable of relating either material properties or material synthesis parameters to material performance quantified as rate constants in microkinetic models. With a robust forward model in place, an inverse model can then be solved to find the material properties that yield the desired performance. To build robust predictive models, Bayesian statistical methods are preferred because they more completely capture the uncertainty in parameter estimates.;To support the discovery informatics system, large amounts of experimental data need to be collected and stored in a database. This work relates the experimental data to a mathematical graph that describes how material move through unit operations and suggests design parameters for a database to store such information. A review of heterogeneous catalyst synthesis techniques is then reviewed as an introduction to the scale of the problem and a transfer function model for material properties and unit operations is proposed.;Bayesian microkinetic modeling is contrasted with traditional least squares optimization for the forward modeling of Epon 825 epoxy curing with 3, aminophenyl-sulfone kinetic data from differential scanning calorimetry. Proper baseline correction procedures are discussed. Five models were fit and it was shown that a ter molecular model, while still inadequate, is most capable of explaining the data. Bayesian parameter estimates are shown to be in line with the least squares estimates because the least squares problem is well posed.;The mechanism of the water gas shift reaction over Pt/Alumina is reviewed and a Bayesian microkinetic model with coverage dependent activation barriers and eight free parameters is fit to experimental data of 20 steady state conversion measurements. It is shown that relying on traditional optimization techniques and parameter point estimates can result in incorrect conclusions because, when uncertainty is taken into account, the activation barrier for the formation of carboxyl overlaps with the activation barriers for possible decomposition routes. It is therefore unclear which step has the largest activation barrier whereas if one chooses the point estimates, one would erroneously conclude that the decomposition by free sites does. A parameter sensitivity study suggests an alternate parameterization that, when fit, predicts a very different surface coverage of hydroxyl groups, but equally well fits the data. Additional experiments are needed to select the correct model.