Understanding and identifying detachment mechanisms of unsteady partial cavitation, successively governed by re-entrant jets and condensation fronts (also referred to as bubbly shocks or condensation shocks), is essential for effective control and utilization of cavitating flows. In this study, a machine learning-based framework is proposed to identify the development stages of partial cavitation in an axisymmetric Venturi by integrating dimensionality reduction with unsupervised clustering, without requiring prior labeling or empirical assumptions. High-speed imaging snapshots of cavitating flow field at σ = 0.56 were analyzed using spectral proper orthogonal decomposition combined with t-distributed stochastic neighbor embedding. This approach reduced the flow field dimensionality from 104 320 to 3 while preserving dominant spatiotemporal features of cavitation evolution. Subsequently, three unsupervised clustering algorithms, density peaks clustering (DPC), K-means (KM), and mean shift (MS), were independently applied to identify distinct cavitation development stages. Among them, the DPC demonstrated superior performance, successfully identifying six cavitation stages: cavitation inception, sheet cavity growth, re-entrant jet development, local sheet cavity shedding, cloud collapse accompanied by condensation front propagation, and rapid retraction of residual cavity. In contrast, both the KM and MS failed to clearly distinguish between the first and sixth stages, and the KM also shows ambiguity in differentiating the fourth and sixth stages. The robustness of DPC was further validated at σ = 0.49 and 0.60, where it accurately captured key characteristics of the propagation of condensation front and its arrival at the throat. Overall, this study presents a robust, data-driven methodology for automated cavitation-stage identification, offering new insights into the dynamics of complex cavitating flows.
Liu et al. (2026) studied this question.