Introduction

Process design involves the entire procedure from raw materials to products, combining multiple unit operations, such as reaction, heat exchange, and separation. In the development of new production processes, the design and performance of virtual plants are simulated. In this study, the objective is to design an efficient ethylene oxide (EO) production process.
EO is an important petrochemical product formed by the oxidation of ethylene. It is converted to various compounds such as ethylene glycol and ethanolamine, and is a starting material for fibers and detergents. EO production is about 9,000 t/y1 in Japan. In the production process, the oxidation of ethylene is accompanied by side reactions:
\({2\ C}_{2}H_{4}+O_{2}\rightarrow{2\ C}_{2}H_{4}O\) (1)
\(C_{2}H_{4}+3O_{2}\rightarrow 2\text{CO}_{2}+{2H}_{2}O\) (2)
\({2\ C}_{2}H_{4}O+5\ O_{2}\rightarrow 4\ \text{CO}_{2}+{4\ H}_{2}O\)(3)
Eq. (1) is the main reaction to produce EO and it requires silver catalysts (Ag2O). Eqs. (2,3) are side reactions that produce carbon dioxide and water. The EO process has a risk of explosion because Eq. (2) has a heat reaction of +1323 kJ/mol at 523 K. Therefore, there is a restriction on reactor inlet compositions and conversions2. As a result, unreacted gas is recycled, and the reactor exhibits nonlinear and complex behaviors. Here, we design the EO production process assuming that only these three reactions proceed.
Much research has been conducted to improve the performance of EO production. Zhou et al. used a reactor model to optimize design variables such as inlet compositions of ethylene, oxygen, and carbon dioxide, as well as reactor temperature and the amount of ethylene dichloride 3. Lahiri et al. optimized design variables, such as inlet compositions and gas flow, using a support vector machine and a genetic algorithm for EO recycling. They verified the results with an actual plant based on the optimized design variables, and reported improved EO production rate and catalyst selectivity4. Yang et al. performed modeling and validations for steady-state, dynamic, and start-up operation conditions with EO recycling. They speeded up the plant start-up with optimized start-up profiles for inlet compositions, temperature, and pressure5. Rahimpour et al. modeled catalyst deactivation using artificial neural networks to determine the optimal amount of ethylene dichloride6. Luo et al. integrated reaction kinetics, a reactor model, and a catalyst deactivation model of an industrial ethylene oxide reactor and compared the results with an actual EO plant7. Nawaz et al. used Aspen Custom Modeler to optimize several design variables, such as reactor temperature and inlet compositions8. Peschel et al. considered interactions between reaction concepts and the entire EO process, and modeled the physical and chemical principles. They reported reduced operating costs and reduced carbon dioxide emissions9. Distillation columns for the removal of carbon dioxide were optimized for several variables, such as temperature and pressure10, 11.
Often, the values of design variables are optimized for each unit, such as a reactor and a distillation column traditionally. Hence, effects between units are not considered, and the optimized values of design variables from these methods are not always accurate when considering the entire process.
We propose a method for optimizing the entire process by considering effects between the units. Therefore, all units must be optimized simultaneously, with a very large number of design variables that need to be optimized. For example, the design variables X for the EO process, such as equipment and operating conditions, have 23 parameters. Assuming there are ten candidates for each X, the total number of combinations would be 1023, which requires a huge number of simulations.
Hence, we propose a process design method based on Bayesian optimization12 where X values that satisfy target values of multiple objective variables Y, such as yield, are searched. Then X will be optimized with a small number of simulations. We verify the effectiveness of the proposed method by simulating an EO plant.