Real-Time Modeling of Gas-Liquid Two-Phase Flow through Machine Learning and Optimization Based on Experimental Data

Abstract

This work aims to enhance the gas-liquid two-phase flow model to improve the accuracy of predicting fission product removal via pool scrubbing, in case of a severe accident at a nuclear power plant. The complexity and computational cost associated with calculating detailed flow parameters under pool scrubbing conditions pose challenges. To address this, development of a real-time simulator was considered, to account for the need for fast calculations to facilitate model refinements. Gas-liquid two-phase flow calculations were performed based on Navier-Stokes equations in particle simulations. The resulting database of particle simulations were used to train a regression forest model. Sensitivity analysis identified key parameters in the trained model, and they were optimized through multi-parametric Bayesian optimization by comparing statistical flow parameters with the experimental data. With such data assimilation, the regression model was refined to output more realistic simulation results. This study highlights the potential of the use of machine learning in the areas of realistic real-time analysis, offering a promising method for improving the prediction capabilities of gas-liquid two-phase flow.