Applying Machine Learning Methods to Forecast Heat Transfer Characteristics in Air-Cooled Single-Row Finned Tube Bundles under Free Convection Conditions

The paper presents the thoughtful experience of developing and applying both classical and advanced machine learning (ML) methods to the regression task aiming at the prediction of heat transfer intensity of air-cooled heat exchangers in the form of horizontal single-row finned tube bundles streamlined by laminar free convective airflow. The initial data for training ML models were tabular datasets obtained in numerical simulations on the problem of changing the transversal tube pitch and the temperature difference between the base tube and the ambient air. The possibilities, quality, and performance of a number of classical and advanced ML methods were studied to forecast with acceptable precision the mean Nusselt number for the bundles with two varying parameters. An overview of the current state of experimental and numerical studies on the problem of tube bundles of a specified type and operating conditions, as well as the application of ML methods in the fields of fluid mechanics and heat transfer, with an emphasis on the physical problem under consideration, is made. The physical prototype and problem setup, as well as the mathematical model and computational aspects of CFD simulations, are described; examples of typical velocity and temperature fields and local heat transfer characteristics are provided. A brief description of ML models, evaluation metrics of quality, computing resources, and program instruments used in the research is given. The results part includes comparison and assessment of the quality of models, presentation, and analysis of predictions of the Nusselt number for the two datasets involved. Novel correlations constructed by the models on the basis of the CFD data are proposed.