Evaluation of Physics-Informed Machine Learning Models for Liquid Entrainment during Reflood Transient Using NRC/PSU RBHT Data

Improved prediction of liquid droplet entrainment for post-dryout reflood heat transfer through the development of physics-based models is of high priority for enhancing the capability of nuclear reactor thermal hydraulic codes. However, due to the very complex and transient two-phase flow behavior during reflood, theoretical modeling of the mass and heat transport processes are extremely difficult. As a result, most liquid entrainment models available today are empirical correlations which are typically found to over-predict the droplet entrainment. The current paper aims at developing a physics-informed machine learning (PIML) model that could deliver more accurate evaluation of the two-phase flow entrainment behavior with improved model reliability and efficiency. Based on the comprehensive experimental data obtained from the NRC/PSU RBHT reflood tests, various pure-ML and PIML models have been developed and assessed. It is found that both pure-ML and PIML models can capture overall entertainment correctly and significantly improve the precision accuracy as compared to the conventional models. Random forest ML architecture typically had better performance than that of artificial neural network (ANN) architecture. In addition, effects of the newly developed PIML models on TRACE reflood transient simulations were also investigated.