PINNs4Drops: Convolutional Feature-Enhanced Physics-Informed Neural Networks for Reconstructing Two-Phase Flows

Two-phase flow phenomena play a key role in many engineering applications, including hydrogen fuel cells, spray cooling techniques and combustion. Specialized techniques like shadowgraphy and particle image velocimetry can reveal gas-liquid interface evolution and internal velocity fields; however, they are largely limited to planar measurements, while flow dynamics are inherently three-dimensional (3D). Deep learning techniques based on convolutional neural networks provide a powerful approach for volumetric reconstruction based on the experimental data by leveraging spatial structure of images and extracting context-rich features. Building on this foundation, Physics-informed neural networks (PINNs) offer a complementary and promising alternative integrating prior knowledge in the form of governing equations into the networks training process. This integration enables accurate predictions even with limited data. By combining the strengths of both approaches, we propose a novel convolutional feature-enhanced PINNs framework, designed for the spatio-temporal reconstruction of two-phase flows from color-coded shadowgraphy images. The proposed approach is first validated on synthetic data generated through direct numerical simulation, demonstrating high spatial accuracy in reconstructing the three-dimensional gas-liquid interface, along with the inferred velocity and pressure fields. Subsequently, we apply this method to interface reconstruction for an impinging droplet using planar experimental data, highlighting the practical applicability and significant potential of the proposed approach to real-world fluid dynamics analysis.