A physics-informed, data-driven framework for estimation and optimization of two-phase pressure drop of refrigerants in mini- and macro channels

At present, there is no universally accepted approach for forecasting the flow boiling heat transfer in the scientific literature. The objective of this study is to predict and optimize the two-phase pressure drop (2ϕΔP) for different refrigerants in both mini and macro-sized channels under various operating conditions. The dataset is amassed from open literature and consists of 1954 points that include seven different refrigerants, including R12345yf, R1234ze, R134A, R410A, R22, R744, and R717 containing heat flux (Q) in the range 0–40 kW/m², hydraulic diameters (Dh) in the range 0.509–14 mm, mass flux in the range 55 < G < 1800 kg/m²s, saturation temperature (Tsat) in the range -40 – 43 °C, and flow qualities (x) in the range 0.001–0959. The first step involves conducting feature engineering operations to select the most significant and impactful features for estimating the considered output. The interpretation by Pearson's correlation highlighted that Dh, Tsat, x, Q, and G were the most impactful params. The Bayesian optimization incorporated with surrogacy was used to determine the required hyper-parameters, which were then deployed to form multi-layer deep neural network (DNN) for estimating 2ϕΔP. A metaheuristic algorithm known as differential evolution (GA) was used to estimate Dh, Tsat, x, Q, and G to minimize the pressure drop. The performance of the established model was determined to be highly accurate with an accuracy of 0.9985. Additionally, an appropriate 2ϕΔP value for each refrigerant was provided to obtain the optimized values of the different input features. The optimized ML (regression) models performed better than the existing semi-empirical pressure drop correlations. This approach can be beneficial for the development and improvement of various engineering systems, particularly those involving the 2ϕΔP of numerous refrigerants.