Heat transfer coefficient prediction AI system for mini-channels based on thermoinformatics technology

In recent years, heat transfer in mini-channels, whose internal diameters are 1–4 mm, has been studied in several applications, such as refrigeration, air conditioning, radiators, heat pipes for electronics, heat radiation in space, and heart-lung machines. In particular, in the areas of refrigeration and air conditioning, several studies have been conducted about the heat transfer in mini-channels due to their ability to considerably improve heat transfer efficiency. In addition, mini-channels make the heat exchanger compact and aid in reducing the production amount of refrigerants, which have a negative influence on the global environment.

Although mini-channels have many advantages as previously described, the prediction of heat transfer coefficients of mini-channels is extremely difficult because the heat transfer area during liquid film-conductive evaporation varies based on the changes in the gas plug, flow condition, and flow direction caused by the flow amount and quality.

As there is big data for heat transfer obtained by various conditions, artificial intelligence (AI) technologies can be used as an alternative method for predicting heat transfer coefficients. We are currently working on various thermal engineering and informatics issues, including multiphase flow, and we refer to this research area as “thermoinformatics,” a new field that combines thermal engineering and informatics. Our previous study demonstrated the ability of deep learning to predict heat transfer coefficients with relatively high accuracy. However, general AI systems do not guarantee that their output will approximately match the correct value. In some cases, they may predict values that are far off, but it is difficult to know this in advance, which is a major practical problem.

Herein, we combine deep learning and Gaussian process regression to predict not only heat transfer coefficients but also their uncertainties based on their variances. Experiments have confirmed that the proposed system can predict with higher accuracy than existing methods. Moreover, it is possible to output the uncertainty for each prediction value, and it was confirmed that the uncertainty value also increases when the prediction error is large. This means that a highly reliable thermoinformatics system has been constructed.