Prediction of subgrid interfacial area in two-phase turbulent flows using convolutional neural network-based architectures

The interfacial area is a very important physical parameter in two-phase flows, which governs the mass, momentum, and energy transfer between the phases. Large-eddy simulations (LES) of two-phase flows employ empirical models for interfacial area to close the effect of unresolved scales. In this talk, we will present data-driven models based on convolutional neural network architectures like autoencoders, generative adversarial networks, and transformers for the prediction of subgrid interfacial area. We will use a fundamental flow, such as stationary two-phase homogeneous isotropic turbulence, to train the models over a wide range of Taylor-scale Reynolds numbers (87-555) and integral-scale Weber numbers (20-2200), including different filter widths with simulation datasets up to 4096^3 grid points. This approach provides a data-driven alternative to regime-limited models; thereby advancing the LES modeling of two-phase turbulent flows for simulations in complex settings.