Systematic development of multilayer-perceptron-based void fraction model

This study presents a multilayer perceptron (MLP) model for approximating the void fraction of a gas–liquid two-phase flow in a smooth horizontal tube. The MLP model was developed using 1618 void fraction measurements collected from 14 data sources associated with refrigerant flow under adiabatic or evaporative heat transfer. The optimal architecture of the MLP model was determined based on a systematic development method that provided the criteria for selecting the optimal network architecture. The proposed method prevented unnecessary increments in model complexity by searching for a compromise between the fitting and generalization abilities of the MLP model. The prediction performance of the MLP model was comprehensively assessed for each data source and vapor-quality range. A comparison of the assessment results with the currently available universal model exhibited a superior overall performance of the MLP model, with a root mean square error of 0.032, mean percentage error of 0.574 %, mean absolute percentage error of 3.691 %, and a coefficient of determination of 0.971 for the entire dataset.