Deep learning strategy for sustainable chemical process

<p dir="ltr">The climatological effect of CO2 has recently prompted the transition of chemical processes towards sustainability, aiming to create a more environmentally friendly and socially responsible industry. However, due to intricate and non-linear behaviour, it is challenging to perform prediction and optimisation in sustainable chemical processes in the real industrial environment. In this work, a hybrid approach, embedding mechanistic insights into the data-driven model, is developed and utilised as the surrogate model, incorporated with a non-dominated sorted genetic algorithm (NSGA-II) to enable a robust prediction and an effective multi-objective optimisation in the sustainable chemical processes. To extend the applicability of this AI-enabled optimisation framework, it is implemented in the adaptive optimisation paradigm, dynamically adapting to real-time changes for maximum effectiveness.</p><p dir="ltr">This thesis is written in an alternative format, where it includes a few coherent and continuous publications from Chapter 2 to 5. In the first chapter, a brief project background is given with an in-depth literature review for the research work, and finally summarised with the main objectives and alternative thesis structure. Chapter 2 is the published review article, envisioning a future smart operation in the sustainable chemical processes when emerging artificial intelligence (AI) as a key to closing the loop of cyber and physical systems, physically facilitated by additive manufacturing (AM). Chapter 3, 4 and 5 focus on the progressive development of the robust prediction and optimisation model empowered by AI in sustainable chemical processes. Chapter 3 presents the development of the simple hybrid model, integrating the kinetic insights from the continuum lumping kinetic (CLK) model to a shallow artificial neural network (ANN) as the surrogate model, coupled with NSGA-II algorithm in the dual-objective optimisation, aiming selectivity and product yield in the hydrocracking process. Chapter 4 demonstrates an improved deep surrogate model by integrating the multiphysics information into a deep neural network (DNN) in the tri-objective optimisation problem, optimising performance, durability, and cost-effectiveness in the PEM fuel cell with graded electrode design. Furthermore, the similar complexity of the AI-enabled optimisation framework introduced in Chapter 4 is further developed in Chapter 5 to perform adaptive optimisation of electrochemical CO2 reduction reaction (eCO2RR). In this case, the AI-enabled optimisation framework is adopted as a smart control system, capable of responding to intermittent changes in renewable energy predicted by Bi-LSTM (Bidirectional long short-term memory) while maintaining high performance in the eCO2RR. Finally, the last chapter summarises the outcome of each chapter, followed by the future work.</p>