A Novel Approach for Reducing Feature Space Dimensionality and Developing a Universal Machine Learning Model for Coated Tubes in Cross-Flow Heat Exchangers

Abstract

In the thermal industry, one common way to transfer heat between hot tubes and cooling fluid is using cross-flow heat exchangers. For heat exchangers, microscale coatings are conventional safeguards for tubes from corrosion and dust accumulation. This study presents the hypothesis that incorporating domain knowledge based on governing equations can be beneficial for developing machine learning models for CFD results, given the available data. Additionally, this work proposes a novel approach for combining variables in heat exchangers and building machine learning models to forecast heat transfer in heat exchangers for turbulent flow. To develop these models, a dataset consisting of nearly 1000 cases was generated by varying different variables. The simulation results obtained from our study confirm that the proposed method would improve the coefficient of determination (R-squared) for trained models in unseen datasets. For the unseen data, the R-squared values for random forest, K-Nearest Neighbors, and support vector regression were determined to be 0.9810, 0.9037, and 0.9754, respectively. These results indicate the effectiveness and utility of our proposed model in predicting heat transfer in various types of heat exchangers.

Keywords

• computational heat transfer
• coating
• feature combination
• machine learning
• heat exchangers