Elucidation of chemical reaction networks through genetic algorithm

• Main Author: Hii, Charles Jun Khiong
• Published: University of Newcastle upon Tyne 2017
• Subjects: 541
• Online Access: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.724706

Obtaining chemical reaction network experimentally is a time consuming and expensive method. It requires a lot of specialised equipment and expertise in order to achieve concrete results. Using data mining method on available quantitative information such as concentration data of chemical species can help build the chemical reaction network faster, cheaper and with less expertise. The aim of this work is to design an automated system to determine the chemical reaction network (CRN) from the concentration data of participating chemical species in an isothermal chemical batch reactor. Evolutionary algorithm ability to evolve optimum results for a non-linear problem is chosen as the method to go forward. Genetic algorithm’s simplicity is modified such that it can be used to model the CRN with just integers. The developed automated system has shown it can elucidate the CRN of two fictitious CRNs requiring only a few a priori information such as initial chemical species concentration and molecular weight of chemical species. Robustness of the automated system is tested multiple times with different level of noise in system and introduction of unmeasured chemical species and uninvolved chemical species. The automated system is also tested against an experimental data from the reaction of trimethyl orthoacetate and allyl alcohol which had shown mixed results. This had prompted for the inclusion of NSGA-II algorithm in the automated system to increase its ability to discover multiple reactions. At the end of the work, a final form of the automated system is presented which can process datasets from different initial conditions and different operating temperature which shows a good performance in elucidating the CRNs. It is concluded that automated system is susceptible to ‘overfitting’ where it designs its CRN structure to fit the measured chemical species but with enough variation in the data, it had shown it is capable of elucidating the true CRN even in the presence of unmeasured chemical species, noise and unrelated chemical species.