Semi-automatic identification of reaction networks using Mixed-Integer Linear Programming (MILP)

Abstract

i Abstract Understanding the processes taking place within a reaction network are of critical importance, particularly for new product development in the chemical processing industries. With more data from new and unknown processes being collected, an efficient method of converting this data into useful information explaining the underlying chemistry would be desirable. Traditionally the method of identification of reaction kinetics was carried out manually using the integral and differential methods. However, when multiple and/or complex reactions are taking place these methods do not compute accurate values. This work takes these manual methods and uses Mixed Integer Linear Programming (MILP) to automate the identification of the stoichiometric and kinetic models of a simulation of a simplified biodiesel production reaction network under a variety of different conditions, and an experimental dataset of the thermal decomposition of 𝛼-pinene. The stoichiometries of the biodiesel production reaction network were identified under a variety of noise (0-5%) and measurement levels (2-75) using MILP using only concentration measurements as inputs. The results indicate that there is a minimum required number of measurements – 25 measurements for this experimental data, for the algorithm to identify stoichiometries. The quality of results also decreases once the amount of noise increases over approximately 2.5%, although it is still possible to find the stoichiometries above that point. The identification of chemical reaction kinetics of the biodiesel production reaction network is also achieved using MILP at a varying noise levels (0-5%) and number of measurement levels (2-75). It has shown that kinetic structures can be identified with only concentration measurements as inputs. As with the stoichiometric identification algorithm, the kinetics identification requires a minimum of 10 measurements to find an accurate model, but the algorithm is most effective when there is at least 30 measurements for this experimental dataset. Similarly once the noise level exceeds approximately 2.5%, the algorithm struggles to reliably identify kinetic models.