A Machine Learning Approach Based on Theoretical Frameworks to Predict Reynolds Stress in Bubbly Flow

This study investigates the prediction of Reynolds stress, a fundamental parameter in turbulent flows, using machine learning methods for both single- and two-phase (bubbly) flows in a vertical pipe. Conventionally the Reynolds stress is known to be difficult to predict owing to its nonlinear characteristics. In this study, particle image velocimetry was conducted under conditions where Reynolds numbers range from 800 (laminar) to 20000 (turbulent) and void fractions vary from 0 up to approximately 2%. To ensure a simple predictive model with minimal computational demands, input features for the machine learning model were selected based on established theoretical frameworks. In single phase flow, the input features were determined through the turbulent-viscosity hypothesis, while for two phase flow, the input features were based on the Sato hypothesis. The trained model successfully predicted Reynolds stress in single-phase turbulent regions and various bubbly flow regimes without bubble coalescence. Additionally, after applying a different approach to dimensionless scaling based on physical meaning, we found that prediction stability improved. This demonstrated that by constructing input features based on physical principles, Reynolds stress can be effectively predicted with a minimal number of input features even in two-phase flows.