Applications of Machine Learning to Predict the Chord Length Distribution of Droplets in Oil-Water Dispersions

作者：Li, Yunchao^*; Liu, Daqian; Liu, Lu

来源：JOM, 2022, 74(9): 3498-3510.

DOI：10.1007/s11837-022-05371-4

• SIZE DISTRIBUTION
• FLOW-THROUGH

The chord length distribution (CLD) of oil-water dispersions were experimentally measured with a focused beam reflectance method (FBRM) probe. With the experimental data, models based on multiple-linear regression (MLG), deep neural network (DNN), random forest (RF), and support vector regression (SVR) have been applied to predict the CLD in oil-water systems and to investigate the connection between shear rate (CUT), rotational speed (RPM), chord length (CL), and the counts of different CLs. Beyond the highest absolute value of correlation coefficients (0.504 for Pearson, 0.813 for Spearman, 0.606 for Kendall), CL influences the counts the most, and they are negatively correlated. The variance and scores of different models were compared, which provided a reference for the application of machine-learning algorithms to the prediction of drop size distribution (DSD) in practical production. From the results, the model based on RF has the highest accuracy (over 0.99) on both the testing and training datasets. To examine our tuned hyperparameters, RF also showed the most stable performance on unseen data, which has 0.9926 R-2 and minimum mean absolute error (MAE) of 0.0159. Our results showed that machine-learning algorithms are stable and precise enough to predict CLD in oil-water dispersions.