A Bayesian learning and data mining approach to reaction system identification: Application to biomass conversion

The growing environmental concern over the use of fossil fuels calls for alternative sources of energy with smaller environmental footprint, and biomass-derived fuels have been extensively investigated as a substitute. In biofuels production, the development of reaction networks and kinetic models is unquestionably a major challenge due to the difficulty in characterizing the reaction products. Therefore, there is a need for a better way to retrieve the information about the reaction from the available experimental data. This study uses a data mining and Bayesian learning approach to estimate the reaction network of the acid and base catalyzed hydrous pyrolysis of hemicellulose from Fourier Transform Infrared (FTIR) spectroscopy. Cluster analysis is used to model the system in terms of lumps and a Bayesian network structure-learning algorithm is then used to device a reaction network. Three Bayesian network structure-learning algorithms were implemented to estimate the reaction network. The results from each were identical, indicating that the model representing the reaction network is most probably in the optimal equivalence space. The model was compared against expert-based reaction models and the agreement is encouraging. A useful aspect of this model is its self-updating capability, i.e., the reaction model can provide a quantitative description of the effect of the change in the operation condition from spectroscopic data. Hence, the model may be used for the real time analysis of the investigated process.