Accurate Prediction of Physical Properties of Different Freons Via Different Machine Learning Models

Abstract

Accurate prediction of refrigerant properties is essential for the design, optimization, and energy efficiency of cooling systems. R-12 and R-22 are among the most widely used refrigerants in industrial and domestic applications. Yet, their thermodynamic behavior is highly nonlinear and difficult to represent with traditional models, particularly under varying operating conditions. These challenges underscore the need for advanced data-driven methods, and machine learning presents a promising solution for enhancing predictive performance. In this research, four machine learning techniques were applied. The models considered were Least-Squares Boosting (LSBoost), Generalized Linear Model (GenLin), Multi-Layer Perceptron Artificial Neural Network (MLP-ANN), and Support Vector Regression (SVR). A dataset collected under controlled laboratory conditions was used for training and testing. The goal was to predict three fundamental refrigerant properties, namely density, velocity, and temperature. These properties were selected because they significantly impact refrigerant mass flow rate, heat transfer efficiency, pressure drop, and overall system performance. The analysis showed that LSBoost consistently achieved the highest predictive accuracy. In density estimation, it achieved an RMSE of 7.69E-04, a Pearson correlation coefficient of 0.9935, a coefficient of determination of 0.9968, a mean absolute percentage error of 5.7847, and a Kling-Gupta efficiency of 0.9736. In contrast, the Generalized Linear Model delivered substantially weaker results. In velocity prediction, LSBoost achieved a Pearson correlation coefficient of 0.9982 and an RMSE of 2.30 x 10(-5), with a coefficient of determination of 0.9978. At the same time, the Multi-Layer Perceptron reached only 0.2132 for the coefficient of determination. For temperature estimation, LSBoost produced an RMSE of 4.3644, a Pearson correlation coefficient of 0.9983, and a coefficient of determination of 0.9992, in contrast to Support Vector Regression, which yielded an RMSE of 92.344. Overall, the study demonstrated that LSBoost was a highly accurate and reliable tool for predicting complex thermodynamic properties of refrigerants, supporting its application in thermodynamic modeling and refrigeration system design. However, the relatively small dataset and simplified input features represented important limitations that may reduce generalizability. Future research should therefore involve larger and more diverse datasets to strengthen the applicability of the proposed approach in real-world engineering contexts.