A Bayesian Learning Approach to Modeling Pseudoreaction Networks for Complex Reacting Systems: Application to the Mild Visbreaking of Bitumen

Abstract

A data-mining and Bayesian learning approach is used to model the reaction network of a low-temperature (150–400 °C) visbreaking process for field upgrading of oil sands bitumen. Obtaining mechanistic and kinetic descriptions for the chemistry involved in this process is a significant challenge because of the compositional complexity of bitumen and the associated analytical challenges. Lumped models based on a preconceived reaction network might be unsatisfactory in describing the key conversion steps of the actual process. Fourier transform infrared spectra of products produced at different operating conditions (temperature and time of processing) of the visbreaking process were collected. Bayesian agglomerative hierarchical cluster analysis was employed to obtain groups of pseudospecies with similar spectroscopic properties. Then, a Bayesian structure-learning algorithm was used to develop the corresponding reaction network. The final reaction network model was compared to the anticipated reaction network of thermal cracking of a model alkyl tricyclic naphthenoaromatic compound, and the agreement was encouraging. The reaction model also indicates that the outcome of thermal processing is the increase in lighter and more aliphatic products, which is consistent with experimental findings. Pseudokinetics were obtained for the reactions between the pseudospecies based on the estimated parameters of the Bayesian network. An attractive feature of the model is that it can be embedded into a process control system to perform real-time online analysis of the reactions both qualitatively and quantitatively.

Chemoinformatic Investigation of the Chemistry of Cellulose and Lignin Derivatives in Hydrous Pyrolysis

“…Figure shows the pseudoconcentrations of the pseudocomponents for the different experimental runs, and it is clear that these concentrations change with different experimental conditions. In a continuously operating process, these concentration signatures could be used with pseudokinetics , derived from the BNs for on-line monitoring of the process.…”

“…While there is a long history of applying chemometric methods in the interpretation and analysis of spectroscopic data, with principal component analysis (PCA) being the most common method, , we focus on methods that aim to deconvolve the spectra into different pseudocomponents or to group elements of the spectra. We refer to our previous work for details of our implementation of these methods: self-modeling multivariate curve resolution (SMCR) and Bayesian hierarchical clustering and only provide short descriptions here.…”

“…Suppose a data allocation K groups all observations into K clusters, and T is the total size of all individual clusters T 1 , T 2 ,...., T K . The random variables are assumed to follow a multinomial distribution with a Dirichlet conjugate prior: Then, the clustering posterior can be represented as Based on this, the individual samples can be grouped into K clusters where the marginal density of the data is f ( x | K ) with a normalization constant larger than zero with no role in clustering (β) …”

“…The main contributions of this work are the validation of our previously developed , methods based on self-modeling multivariate curve resolution (SMCR) and Bayesian hierarchical clustering (BHC) for pseudocomponent identification on the better-understood chemistry of the pyrolysis of model compounds representative of biomass (we have previously demonstrated the application of these methods on the partial upgrading of bitumen), the extension of the SMCR method to develop candidate reaction networks based on its pseudocomponents, and the development of a methodology to hypothesize reaction networks (using the Bayesian network (BN) methodology based on SMCR and BHC-generated pseudocomponents) by fusing data from multiple sources. We demonstrate the chemical feasibility of the data-driven hypotheses generated for reaction networks and also highlight the impact of data fusion/combination on modifying these hypotheses.…”

Data‐driven hypotheses of reaction networks for thermochemical conversion of a physical mixture of levoglucosan and 2‐phenoxyethyl benzene

“…The methods used in this work have been explained in detail in our previous work [1,9,10] ; consequently, we provide only brief descriptions of the specific techniques used here. Note that the headings of the subsections provide both the names of the methods and their purpose.…”