Determination of bubble sizes in bubble column reactors with machine learning regression methods

• Published in: Chemical engineering research & design 2020-11, Vol.163, p.47-57
• Main Authors: Theßeling, Christin, Grünewald, Marcus, Biessey, Philip
• Format: Article
• Language: English
• Subjects: Algorithms; Bubble columns; Bubbles; Deionization; Diameters; Finite element method; Gas–liquid flow; Labels; LASSO regression; Machine learning; Measurement; Multiphase flow measurement; Regression analysis; Regression models; Regression tree algorithm; Spatial resolution; Studies; Supervised learning; Wire mesh sensor
• ISSN: 0263-8762; 1744-3563
• DOI: 10.1016/j.cherd.2020.08.020

•Two models were developed based on wire mesh sensor measurement data.•A systematic approach for derivation of labels from experimental data is introduced.•Dimensionality reduction is crucial for model accuracy and estimating label impact.•LASSO model performs similarly to conventional approach for bubble size derivation.•Regression tree algorithm reduces prediction error of small bubbles significantly. In this study, two machine learning based regression models are developed to predict diameters of single bubbles in a bubble column reactor based on wire-mesh sensor (WMS) measurement. Both Least Absolute Shrinkage and Selection Operator (LASSO) regression and a regression tree algorithm are used to predict bubble diameter with supervised learning techniques. Measurements are carried out in a DN150 column filled with deionized water and air as the continuous phase while WMS passage of single bubbles is investigated. A novel method for definition of different labels characterizing the passing bubble is introduced. Based on the defined labels, Machine Learning regression models are developed to predict bubble sizes. Methods for dimensionality reduction are applied, allowing for an investigation of each labels influence on model prediction quality. Both regression models perform similar or better than well-established approaches to calculate bubble diameter based on WMS measurement. As a highlight, it is shown that bubble diameters even below the sensor’s spatial resolution can be predicted with an accuracy of ±13% using the regression tree model, which is about 1/3 of the conventionally assumed measurement uncertainty at bubble diameters below the sensor’s spatial resolution.