Physics-Constrained Machine Learning for Two-Phase Flow Simulation Using Deep Learning-Based Closure Relation

摘要

1D area-averaged two-phase models have been widely used in system-level thermal-fluid codes due to its fast-running feature. The family of these two-phase models includes the two-fluid model (TFM) as well as two-phase mixture models (TMM). While details vary, the closure relations (CR) or sub-grid-scale (SGS) physics models are essential and dominate the predictability when we model the vapor-liquid pipe flow using the partial differential equations (PDE) for mass, momentum, and energy. The more complex the two-phase model is, the more closure relations are required. Classically, building SGS physics models requires analytical models, and then the model is calibrated using the relevant experiment data. In the meanwhile, the two-phase flow simulation involves the flow regime transition and this requires data from different measurements. The usability of traditional CR relies on the flow regimes map. However, this increases the difficulty for the scaling analysis since each experiment has its own applicable domains and uncertainties. In this work, we explore the hypothesis that the data-driven approach by using machine learning (ML) methodologies can discover the underlying correlation behind the data and enable universal models across flow regimes.