Multi-objective optimisation with hybrid machine learning strategy for complex catalytic processes

Abstract

Catalytic chemical processes such as hydrocracking, gasification and pyrolysis play a vital role in the renewable energy and net zero transition. Due to the complex and non-linear behaviours during operation, catalytic chemical processes require a powerful modelling tool for prediction and optimisation for smart operation, speedy green process routes discovery and rapid process design. However, challenges remain due to the lack of an effective modelling and optimisation toolbox, which requires not only a precise analysis but also a fast optimisation. Here, we propose a hybrid machine learning strategy by embedding the physics-based continuum lumping kinetic model into the data-driven artificial neural network framework. This hybrid model is adopted as the surrogate model in the multi-objective optimisation and demonstrated in the benchmarking of a hydrocracking process. The results show that the novel hybrid surrogate model exhibits the mean square error less than 0.01 by comparing with the physics-based simulation results. This well-trained hybrid model was then integrated with non-dominated-sort genetic algorithm (NSGA-II) as the surrogate model to evaluate and optimise the yield and selectivity of the hydrocracking process. The Pareto front from the multi-objective optimisation was able to identify the trade-off curve between the objective functions which is essential for the decision-making during process design. Our work indicates that adopting the hybrid machine learning strategy as the surrogate model in the multi-objective optimisation is a promising approach in various complex catalytic chemical processes to enable an accurate computation as well as a rapid optimisation.