Machine learning-aided characterization of microbubbles for venturi bubble generator

• Published in: Chemical engineering journal (Lausanne, Switzerland : 1996) Switzerland : 1996), 2023-06, Vol.465, p.142763, Article 142763
• Main Authors: Ruan, Jian, Zhou, Hang, Ding, Zhiming, Zhang, Yaheng, Zhao, Luhaibo, Zhang, Jie, Tang, Zhiyong
• Format: Article
• Language: English
• Subjects: Machine learning; Microbubbles; Multi-dimensional dataset; Sauter mean diameter; Venturi bubble generator
• ISSN: 1385-8947; 1873-3212
• DOI: 10.1016/j.cej.2023.142763

[Display omitted] •Development of a method for characterization of microbubbles for venturi tube based on machine learning.•Full factorial design of experiment is performed to obtain reliable data for training three ML models.•Three ML models exhibit excellent predictability on the Sauter mean diameter.•The throat size shows the greatest influence on the Sauter mean diameter.•The ML models outperform mathematical model for predicting the Sauter mean diameter. The characterization of microbubbles for venturi tube is important for the associated industrial applications, but still challenging due to the coupling effects of numerous operating factors. Here, we report a machine learning (ML)-aided approach for predicting the characteristics of microbubbles generated by venturi tube. Full factorial design of experiments (DOE) was first carried out, followed by the image post-processing to obtain multi-dimensional dataset. After data cleaning, MLP (Multi-Layer Perception), random forest (RF) and Catboost models were trained to correlate the Sauter mean diameter (ds) to five operating features, namely, throat-to-outlet ratio β, divergent angle θ, gas-to-liquid ratio α, gas Reynolds number Reg and liquid Reynolds number Rel. All three ML models provide excellent predictability on ds, while the Catboost model displays the best extrapolation performance in three investigated scenarios. Internal importance analysis shows that the throat size and Reg play the greatest and least influence on ds, respectively. We also explored the mathematical fitting approach based on obtained experimental dataset. The results show that ML models deliver improved predictive performance over mathematical model, but the latter provides better mechanistic interpretability. This work demonstrates the great potential of ML in the gas–liquid multiphase flow.