Developing a universal model for predicting condensation heat transfer coefficients remains challenging,particularly for steam–non-condensable gas mixtures,owing to the intricate nonlinear interactions between multip...Developing a universal model for predicting condensation heat transfer coefficients remains challenging,particularly for steam–non-condensable gas mixtures,owing to the intricate nonlinear interactions between multiphase flow,heat,and mass transfer phenomena.Data-driven machine learning(ML)shows promise in efficiently and accurately predicting condensation heat transfer coefficients.Research has employed various ML methods—multilayer perceptron neural networks,convolutional-neural-network–based DenseNet,backpropagation neural networks,etc.—to investigate steam condensation with non-condensable gases.However,these exhibit limited extrapolation ability and heavily rely on data quantity owing to their black-box nature.This study proposes a physics-informed ML model that combines physical constraints derived from the modified Nusselt model with conventional data-driven ML techniques.The model's predictive performance is evaluated using a comprehensive database(879 datapoints from 13 studies).A physics-constrained and eight data-driven ML methods are assessed.The results reveal that the physics-constrained approach combined with XGBoost significantly outperforms conventional ML methods on extrapolation datasets(199 datapoints from 3 studies),achieving a mean absolute percentage error of 11.22%,which is approximately half that of the best-performing fully data-driven model at 21.63%.The model demonstrates consistent and reliable performance across diverse datasets,making it an effective tool for predicting heat transfer coefficients in steam–non-condensable gas mixtures.By deepening the understanding of the underlying physical processes,the proposed model supports the development of precise and efficient engineering solutions for condensation heat transfer.展开更多
基金supported by the National Research Foundation of Korea(NRF)grant funded by the Korea government(MSIT)(No.RS-2024–00353227 and RS-2024–00411577).
文摘Developing a universal model for predicting condensation heat transfer coefficients remains challenging,particularly for steam–non-condensable gas mixtures,owing to the intricate nonlinear interactions between multiphase flow,heat,and mass transfer phenomena.Data-driven machine learning(ML)shows promise in efficiently and accurately predicting condensation heat transfer coefficients.Research has employed various ML methods—multilayer perceptron neural networks,convolutional-neural-network–based DenseNet,backpropagation neural networks,etc.—to investigate steam condensation with non-condensable gases.However,these exhibit limited extrapolation ability and heavily rely on data quantity owing to their black-box nature.This study proposes a physics-informed ML model that combines physical constraints derived from the modified Nusselt model with conventional data-driven ML techniques.The model's predictive performance is evaluated using a comprehensive database(879 datapoints from 13 studies).A physics-constrained and eight data-driven ML methods are assessed.The results reveal that the physics-constrained approach combined with XGBoost significantly outperforms conventional ML methods on extrapolation datasets(199 datapoints from 3 studies),achieving a mean absolute percentage error of 11.22%,which is approximately half that of the best-performing fully data-driven model at 21.63%.The model demonstrates consistent and reliable performance across diverse datasets,making it an effective tool for predicting heat transfer coefficients in steam–non-condensable gas mixtures.By deepening the understanding of the underlying physical processes,the proposed model supports the development of precise and efficient engineering solutions for condensation heat transfer.
