A hybrid physics-based machine learning framework for oil retention prediction: A comparative analysis of modeling approaches in refrigerant vapor lines.

With the deployment of variable-speed compressors in unitary air conditioning (AC) systems as well as the future implementation of newer HFO refrigerants, there is a need to upgrade line sizing guidelines to account for the effects of oil retention (OR). These new guidelines can be facilitated through developing and applying a predictive model for OR in vapor lines for commonly used refrigerant-lubricant combinations in the heating, ventilation, air conditioning, and refrigeration (HVAC&R) industry. This work aims to evaluate the prediction accuracy of machine learning (ML), physics-based (PB), and hybrid models trained using OR data obtained for horizontal and vertical lines of different diameters (11, 17, and 20 mm). The data include different refrigerants (R134a, R410A, R32, and R1234ze(E)) used with POE32 lubricant under various flow conditions. The ML models were trained and tested using data obtained from over 230 experimental tests. The input parameters for each model were refrigerant conditions (type, temperature, pressure, and mass flow rate), pipeline dimensions and orientation, injected oil mass flow rate, and oil viscosity, with OR as the predicted output. Various model types were investigated and compared, including a purely physics-based (PB) model, two standalone ML models, and two physics-based machine learning-aided (PBMLA) algorithms. A sensitivity analysis was performed to assess the effect of input parameters on the prediction errors. In addition, an extrapolation study was conducted using different refrigerants and various oil grades and types to evaluate the models’ ability to predict OR with acceptable accuracy. The results showed that the standalone ML algorithms exhibited lower accuracy in predicting OR compared to the PB model. Furthermore, the PBMLA models demonstrated a modest improvement in OR prediction accuracy over the purely PB model (improved R2 by up to 10.8 %, and reduced RMSE by up to 4.5 %). Moreover, the parametric analysis revealed that the PBMLA models could address variations of the feature inputs relatively better than the PB model, leading to their higher prediction accuracy compared to all other models. Finally, the extrapolation analysis showed that both the PB and PBMLA models were not limited to specific oil types or grades, suggesting their potential for general applicability in OR prediction and system design.