Hybrid-PINNs approach for predicting high-fidelity flow and heat transfer in printed circuit heat exchangers of sodium-cooled fast reactors

The heat exchanger is the key for connecting the primary and secondary circuits in sodium-cooled fast reactors (SFR), and thermal-hydraulic characteristics estimation is vital for design and safety analysis of SFR. While conventional computational fluid dynamics (CFD) methods require expensive computational cost and physics-informed neural-networks (PINNs) depend on well-developed data, but only limited or no high-fidelity data can be obtainable in real SFR. This study presents effective hybrid-PINNs (h-PINNs) to predict the high-fidelity flow and temperature distributions within printed circuit heat exchanger (PCHE) flow channels from low-fidelity data. The h-PINNs approach primarily consist of three deep neural-networks (DNNs): the first is a data-driven DNN trained to establish the relationship between input coordinates and output low-fidelity data; the second, also data-driven, investigates the nonlinear correlation between low-fidelity and high-fidelity data; and final DNN incorporates physics constraints to refine high-fidelity data generated by second DNN. The performance of presented h-PINNs is evaluated by numerical examples of sodium flow in PCHE flow channels with differently shaped fins. The h-PINNs accurately estimate the velocity and temperature distributions, achieving R2 indicators of over 97.54 % with a few high-fidelity data and 95.54 % without any high-fidelity data. The proposed approach can potentially apply to other heat transfer prediction challenges associated with energy conversion equipments.