(476c) Optimization of Fischer-Tropsch Microchannel Reactor Using Computational Fluid Dynamics and Transfer-Learned Bayesian Optimization.

Fischer-Tropsch(F-T) synthesis is a highly exothermic reaction that converts syngas into long chain hydrocarbons. This reaction cause catalyst deactivation such as sintering and fouling in catalysts due to a high temperature gradient [1]. The microchannel reactor is suitable for F-T synthesis because of its heat transfer efficiency, but high temperature gradient can be occurred. The method of discrete dilution is one method to prevent high temperature gradient which is to divide the reactor into several zones so that a different amount of catalyst is loaded separately [2]. Changes in the method of discrete dilution affect production of long hydrocarbons. Hence, optimization of the loading amount of catalyst for each zone is necessary for productivity of long chain hydrocarbon and preventing catalyst deactivation.

Computational fluid dynamics (CFD) is widely used to analyze temperature profile of F-T synthesis in microchannel [2, 3]. Bayesian optimization (BO) is an effective tool for CFD- based optimization in that it purses the maximum efficiency through a solid theoretical background [4]. However, BO is a cold-start approach, which requires numerous function evaluations until it gives a meaningful result [5, 6]. Transfer learning, a method of using knowledge from a previous task to a new task, which improves performance when modeling a new task. Therefore, integration of BO and transfer-learning for CFD-based optimization reduce the number of CFD simulation runs, and thus alleviating the computational requirement.

In this study, we propose the integration of BO and transfer-learning to optimize the F-T microchannel reactor with a length of 600 mm. In BO, CFD simulation is replaced by surrogate model. We chose the gaussian process as a surrogate model which is suitable for uncertainty estimation when dealing with stochastic data. In this study, the optimization variable is packing length and loading amount of catalyst. The objective function of optimization is selected to decrease the difference between inlet and peak temperature inside the reactor, and increase the productivity of long chain hydrocarbon. First, suppose that the historical data of F-T reactor is presented. After that, the optimization of the microchannel reactor with the inlet temperature changed was set as the target task. The transfer learning is utilized to obtain the gaussian process, which is the initial surrogate model. In the target task optimization, the mean and covariance functions are calculated using source task data as initial values to construct the surrogate model for BO. Since the surrogate model of the source task and the target task are different, it is assumed that there is uncertainty in the source task data when used in the target task optimization. BO used this surrogate model to select the next point to obtain from CFD simulation, then update the surrogate model. The influence of the source task data is adjusted through uncertainty calculated using target task data so that a surrogate model suitable for the target task function is constructed. The proposed model reduces the number of iterations to obtain an optimal value of optimization with CFD model than that of the traditional BO.

References

[1]. Tsakoumis, Nikolaos E., et al. "Deactivation of cobalt based Fischer–Tropsch catalysts: a review." Catalysis Today 154.3-4 (2010): 162-182.

[2]. Na, Jonggeol, et al. "Multi-objective optimization of microchannel reactor for Fischer-Tropsch synthesis using computational fluid dynamics and genetic algorithm." Chemical Engineering Journal 313 (2017): 1521-1534.

[3]. Deshmukh, Soumitra R., et al. "Scale-up of microchannel reactors for Fischer− Tropsch synthesis." Industrial & engineering chemistry research 49.21 (2010): 10883-10888.

[4]. Park, Seongeon, et al. "Multi-objective Bayesian optimization of chemical reactor design using computational fluid dynamics." Computers & Chemical Engineering 119 (2018): 25-37.

[5]. Joy, Tinu Theckel, et al. "A flexible transfer learning framework for Bayesian optimization with convergence guarantee." Expert Systems with Applications 115 (2019): 656-672.

[6]. Chuang, Yao-Chen, et al. "Transfer learning for efficient meta-modeling of process simulations." Chemical Engineering Research and Design 138 (2018): 546-553.