Machine learning-based cavitation stage identification in an axisymmetric Venturi under bubbly shock dominance

Understanding the bubbly shock mechanism in partially cavitating flows is vital for both controlling and harnessing cavitation. Venturis, due to their geometric confinement, provide an ideal platform for investigating this phenomenon. This study introduces an automatic identification method for cavitation stages in an axisymmetric Venturi dominated by bubbly shock, combining unsupervised clustering with supervised classification algorithms, requiring no prior knowledge. Using principal component analysis on high-speed photography data at σ = 0.30, three distinct cavitation stages are identified through a density peaks clustering algorithm combined with three representative classifiers: support vector machine (SVM), random forest (RF), and K-nearest neighbor (KNN). Stage I begins with the initial growth of the sheet cavity and ends when pressure waves first impact it, while stage II starts upon this first impact and ends when the sheet cavity first detaches. Stage III follows stage II. To ensure classification accuracy, two additional features, i.e., the difference and slope of the average gray value over a short interval, are introduced. The classification accuracies for SVM, RF, and KNN at σ = 0.30 are 99.3%, 98.8%, and 98.7%, respectively. In addition, SVM demonstrates superior generalization when tested at σ = 0.27 and 0.37, whereas RF and KNN show relatively weaker adaptability. This study presents a robust, data-driven approach for automated cavitation stage identification and offers valuable insight for analyzing and predicting complex cavitating flow structures.