Résumé : The coherent dynamics of bubble clusters in liquid are of fundamental and industrial importance and are elusive due to the complex interactions of disordered bubble oscillations. Here we introduce and demonstrate unsupervised learning of the coherent physics by combining theory and principal component analysis. From data, the method extracts and quantifies coherent dynamical features based on their energy. We analyze simulation data sets of disordered clusters under harmonic excitation. Results suggest that the coherence is lowered by polydispersity and nonlinearity but in cavitating regimes underlying correlations can be isolated in a single cohererent mode characterized by mean-field interactions, regardless of the degree of disorders. Our study provides a valuable tool and a guidance for future studies on cavitation and nucleation in theory, simulation, and experiments.