Machine learning-based efficient multi-layered precooler design approach for supercritical CO2 cycle

In the current study, a multi-layered approach, combining 3D Reynolds Averaged Navier–Stokes (RANS) model, Artificial Neural Network (ANN), and in-house precooler design and analysis code (PDAC), is developed to investigate the influence of various parameters on the design of the precooler used in sCO2 cycle. The validated 3D-RANS model is utilized to compute thermal and hydraulic characteristics of a zigzag channeled PCHEs operating under the precooler's conditions. The generated CFD data is used to develop empirical correlations and train a machine learning (ML) model based on Artificial Neural Network (ANN). Finally, both trained ML model and empirical correlation are utilized in the PDAC to design and analyze the precooler under different design conditions. Regression data for the precooler, operating in the pseudocritical zone, show that the prediction accuracy of the developed Nusselt number and pressure drop correlation is relatively poor. At the same time, the trained Artificial Neural Network (ANN) can estimate 99% of the data with 90% confidence in the same operating regime. Further, precooler design and analysis code (PDAC) based on the trained ML model can accurately mimic the CFD prediction at a significantly reduced computational time. It is found that pressure losses on the sCO₂-side and water pumping power are exceptionally sensitive to desired exit temperature and channel flow rate. Exit temperature lower than the pseudo critical temperature and higher channel flow rate result in pinch point location well inside the precooler, resulting in a massive rise in the heat exchanger's size and pressure losses. The pinch point can be avoided by choosing the channel mass flow rate ratio of the hot and cold-sides fluid comparable with their ratio of specific heat capacities corresponding to pseudocritical temperature.