Using Machine-Learning-Aided Computational Fluid Dynamics to Facilitate Design of Experiments

Abstract

The design of novel reactors and chemical processes requires an understanding of the fundamental chemical-physical processes at small spatial and temporal scales and a systematic scale-up of these studies to investigate how the process will perform at industrial scales. The financial and temporal costs of these studies can be significant. The use of statistical machine-learning-based methods can significantly reduce these costs. The use of the design of experimental methods can help design an experimental plan that efficiently explores the design space using the fewest experiments possible. Computational methods such as computational fluid dynamics (CFD) are effective tools for detailed studies of small-scale physics and are critical aids to facilitate and understand physical experiments. However, CFD methods can also be time-consuming, often requiring hours or days of time on supercomputers. In this research, we investigate the combination of machine learning with reducing 3D CFD simulation to 2D by exploiting axial symmetry to facilitate the design of experiments. Focusing on a 3D carbon dioxide (CO₂) capture reactor as an example, we demonstrate how machine learning and CFD can help facilitate modeling and design optimization. A 2D CFD is used to simulate the chemical–physical processes in the reactor and is then coupled with machine learning to develop a less computationally expensive model to accurately predict CO₂ adsorption. The learned model can be used to optimize the design of the reactor. This paper demonstrates the decrease in temporal and financial costs of designing industrial-scale chemical processes by combining reducing the CFD dimension and machine learning. Equally importantly, this research demonstrates the significance of selecting a proper machine-learning algorithm for different tasks by comparing the performances of different machine-learning algorithms.