Hybrid modelling of heat transfer systems: Combining physics-based and data-driven approaches for improved prediction and extrapolation

Hybrid machine learning-assisted modelling techniques have gained increasing attention recently in many engineering fields. This is due to the challenges associated with pure first-principles and data-driven models, as the former requires deep phenomenological understanding and might become infeasible to describe a complex system with, and the latter needs extensive high-quality data and, more importantly, extrapolates poorly compared to its first principles counterparts. The integration of the two techniques in a framework will result in an integrated approach that benefits from the two realms by strengthening extrapolation capabilities, higher prediction accuracy, and less data demanding and more data-efficient. In this study, a systematic hybrid modelling framework is developed, allowing for the integration of mechanistic models and machine learning algorithms in parallel and series for modelling heat transfer systems to predict a desired target variable, as long as the system is not of a dynamic nature. The framework is developed according to a previous study that enabled the use of machine learning models for such systems. The application of the hybrid modelling framework in this study is demonstrated on the prediction of the condensation heat transfer coefficient in a microfin tube. A laboratory-scale dataset of 5708 datapoints is used for the validation of the developed framework. The validation of the model has been carried out in two different scenarios, both assessing the general prediction and extrapolation capabilities of the developed models in comparison with pure mechanistic and pure machine learning models. The hybrid models, series and parallel, outperform the mechanistic model by approximately 60% more accurate predictions and the machine learning model by 25%, while interpolating. More importantly, while extrapolating, the hybrid models showed approximately 50% more accurate predictions compared to pure machine learning and 27% more accurate compared to the mechanistic model.