From Images to Hidden Flow Fields: CFD-Guided Machine Learning for Two-Phase Flow Reconstruction

Two-phase flows are governed by dynamic gas-liquid interfaces whose shape, motion and internal flow topology control processes such as spray cooling, coating, inkjet printing and fuel-cell operation. Conventional optical diagnostics mainly deliver two-dimensional projections of inherently three-dimensional interface dynamics, and measurements of internal droplet flow are hampered by refraction at the curved interface. This talk presents a framework that combines experimental imaging, CFD and physics-informed machine learning to reconstruct hidden three-dimensional flow information from image sequences. Glare-point shadowgraphy with color-coded lateral illumination encodes local interface orientation and out-of-plane deformation. Phase-field simulations of droplet impingement combined with physics-based rendering generate synthetic image sequences with known phase, velocity and pressure fields, which are used to train PINNs4Drops, a video-conditioned physics-informed neural network framework constrained by CFD data and the governing equations of incompressible two-phase flow. VoF-based PINNs improve out-of-plane reconstruction, temporal consistency, volume conservation and interface curvature compared with a purely data-driven baseline, and video-conditioned PINNs infer continuous 3D velocity and pressure fields from image sequences with average validation errors of about 6-9%. ... mehrApplied to real droplet-impact experiments, the method recovers physically plausible hidden flow topology from a single optical view.