Title:Classifying acoustic cavitation with machine learning trained on multiple physical models

Abstract:Acoustic cavitation threshold charts are used to map between acoustic parameters (mainly intensity and frequency) and different regimes of acoustic cavitation. The two main regimes are transient cavitation, where a bubble collapses, and stable cavitation, where a bubble undergoes periodic oscillations without collapse. The cavitation charts strongly depend on the physical model used to compute the bubble dynamics and the algorithm for classifying the cavitation threshold. The differences between modeling approaches become especially noticeable for resonant bubbles and when sonication parameters result in large-amplitude oscillations. This paper proposes a machine learning approach that integrates three physical models, i.e., the Rayleigh-Plesset, Keller-Miksis and Gilmore equations, and multiple cavitation classification techniques. Specifically, we classify the cavitation regimes based on the maximum radius, the acoustic Mach number, the kurtosis factor of acoustic emissions, and the Flynn criterion on the inertial and pressure functions. Four machine learning strategies were developed to predict the likelihood of the transient and stable cavitation, using equally weighted contributions from classification techniques. By solving the differential equations for bubble dynamics across a range of sonication and material parameters and applying cross-validation on held-out test data, our framework demonstrates high predictive accuracy for cavitation regimes. This physics-informed machine-learning approach offers probabilistic insights into cavitation likelihood, combining diverse physical models and classification strategies, each contributing different levels of physical rigor and interpretability.