Sample-efficient and generalizable velocity and pressure field reconstruction from three-dimensional two-phase flow interfaces

Abstract

Accurate reconstruction of velocity and pressure fields from gas-liquid interfaces is essential for quantifying two-phase flow dynamics and enabling real-time flow visualization and diagnostics. This study presents a Transformer-based neural operator model pretrained on large-scale two-dimensional interface data and fine-tuned using a limited set of three-dimensional interfaces, enabling mapping gas-liquid interfaces and corresponding governing dimensionless numbers to velocity and pressure fields based on rising bubble simulations generated by the lattice Boltzmann method. Reconstruction errors are systematically evaluated in parameter space as well as in spatial and temporal dimensions. Then the pretrained model is further applied to data-scarce three-dimensional cases serving as a sample-efficient approach to reconstruct high-dimensional data. The results suggest that initialization from a pretrained model always outperforms training from scratch, and that fine-tuning all layers yields the best performance, as the relatively small number of model parameters helps prevent overfitting. These findings demonstrate that model performance in data-limited scenarios can be effectively enhanced through pretraining in the diverse and fast-generated low-dimensional sample space, highlighting a viable pathway toward data-efficient modelling of multiphase flows.