Abstract This paper addresses the critical gap in predicting the heat transfer coefficient during flow boiling in enhanced tubes, where the use of conventional correlations and predictive methods developed for smooth surfaces do not usually provide satisfactory results. For such purpose, a comprehensive database was collected from existing literature, including a wide range of operating conditions and enhanced tube geometries from several independent sources. The dataset includes mass flow rates spanning from 50 to 1000 kg/m2s, vapor qualities from the onset of boiling (x=0.0) to the dry-out occurrence and beyond (x=0.99), reduced pressures from 0.05 to 0.80, and tube diameters (measured up to the fin tip) from 0.7 to 11.9 mm, for a total amount of approximately 3000 data points. Existing flow boiling heat transfer coefficient predictive methods for enhanced tubes were implemented and tested with the present dataset, proving a limited accuracy for most of them mainly in case of testing beyond the specific parameter ranges they were developed for. Extrapolation frequently resulted in statistically poor or even non-physical outcomes. Several artificial neural network models were then developed, according to sensitivity analysis approach to look for potential input parameters and network structures. Specifically, two approaches were employed: a standard neural network model and a correlated informed neural network (CINN), integrating physical correlations into the network’s architecture, thus informing the model with physical principles that govern the heat transfer process. Despite a lower overall accuracy, the correlated informed neural network demonstrated superior reliability than standard one, resulting in an instrument to improve the accuracy of existing correlations.
Read full abstract