14 / 2024-05-29 11:22:35

Artificial intelligence application to predict departure from bubbly flow regime in two-phase flow analysis

Machine learning,Support vector machine,K-Nearest neighbor classifier,Bubbly,Shadowgraph

Two-phase flows are modeled based on their flow regime. Therefore, the transition from one regime to another is critical to a reliable analysis of the flow parameters. Artificial intelligence capabilities have been increasingly applied in different aspects of multiphase flow research recently. Machine learning models such as Artificial Neural Networks, Convolutional Neural Networks, and Random Forests have been applied to predict and analyze two-phase flow patterns, and heat transfer characteristics.

The visual vocabulary can be extracted from the images in the form of a histogram is a representation of different regimes. The Support Vector Machine and K-Nearst Neighbors (KNN) will be used to predict departure from bubbly flow in images acquired in shadowgraph experiments. Applying machine learning models requires a dataset to train and test the models' performance. The training dataset will include flow images from channels, rod bundles, and rising bubbles in a pool. The models will be trained with two datasets of images that are bubbly flow regimes and those that are not bubbly flow regimes. The model will then be tested on images not involved in the training dataset and the open literature. The two categories will allow the analysis to build robustness in testing when a new set of images contains images that are not bubbly flow regimes or vice-versa.

The performance of the models will be evaluated through the confusion matrix and further by how well they perform with images not included in the training or validation dataset. The success of the validation stage is measured by the confusion matrix that summarizes the performance of the models. The trained model will predict bubbly flow regimes from others. These were three distinct regimes observed in the dataset. The model performance in predicting bubbly flow in a random set of data is expected to be above 97.5% accurate.