Data-driven modeling of interface and topological changes in multiphase flow

Multiphase flows pose a persistent challenge in computational fluid dynamics due to the difficulty of accurately tracking interfaces and capturing topological changes such as coalescence and breakup. Traditional solvers—particularly immersed boundary methods—devote over 60% of runtime to resolving phase boundaries and associated dynamics. This cost increases dramatically with finer resolution or more complex interface interactions. In this work, we investigate a machine learning–augmented framework to accelerate multiphase flow simulations by predicting interface evolution and topological transitions. Using data generated from high-fidelity immersed boundary simulations, we explore both transformer-based models and physics-informed neural networks (PINNs) to learn the spatiotemporal evolution of phase interfaces. These models aim to approximate interface behavior and particle trajectories while preserving the underlying physics. This approach offers a path toward reducing computational cost while enabling scalable, physics-aware simulation of complex multiphase dynamics.