Machine learning-based mapping of interfacial dynamics in confined multiphase microflows

Forecasting and classifying flow regimes in confined multiphase flows remain fundamental challenges in thermofluids due to the complex, nonlinear interactions among viscous, inertial, and capillary forces. This study develops an integrated experimental and data-driven analytical framework to map these regimes objectively and analyse interfacial dynamics in a microfluidic channel. High-speed imaging combined with unsupervised learning was used to identify key shape descriptors (Aspect Ratio, Tracer Deformation Index) and essential dimensionless numbers (Reynolds number Re, Capillary number Ca, Weber number We ) to build a predictive feature space. K-means clustering (k=3) identified plug, slug, and parallel flow regimes, with validation indicated by a high silhouette score of 0.68 and by Principal Component Analysis (PCA), which demonstrated clear separation between geometry-dominated and force-dominated contributions to regime classification. Correlation analysis revealed statistically significant associations between deformation and the Capillary number (r = +0.48) and a strong inverse association with plug velocity (r = −0.90), reflecting a confinement- and residence-time-dependent interfacial response rather than causal decoupling of forces. While the present study uses offline image processing, the use of computationally efficient geometric descriptors suggests potential for future real-time monitoring and control of multiphase microflows. The methodology is broadly applicable, though quantitative results are specific to the investigated microchannel geometry and water–silicone oil system.