Transfer learning-driven thermal characterization of heat exchangers in frosting conditions for data-efficient prediction

Abstract

To predict the thermal performance of heat exchangers under frosting conditions, existing empirical correlation-based approaches are time- and cost-intensive because large datasets must be newly collected whenever geometric modifications are introduced into the heat exchanger design. Although data-driven models based on artificial intelligence have attracted attention due to their ability to estimate heat exchanger performance accurately and to extend the prediction range, such models still require substantial amounts of data. Moreover, performance prediction under frosting, which is characterized by time-dependent degradation and pronounced domain shifts across geometries and operating conditions, remains underexplored. In this study, we propose a data-efficient method for predicting the thermal performance of heat exchangers under frosting conditions. Specifically, we employ a transfer learning framework that leverages knowledge from a pre-trained model developed for a conventional fin-tube heat exchanger. When adapted to a fin-tube heat exchanger with a different geometric configuration, the proposed approach achieves high prediction accuracy (coefficient of determination, R2 > 0.95) while using 50% less training data than a deep learning model. As a proof of concept under large domain shifts, the method is further evaluated on an additional target domain with different geometry and operating conditions. This study represents an effort to broaden the application of artificial intelligence to heat exchangers by enabling data-efficient and accurate prediction of thermal performance under diverse frosting conditions and geometric configurations.