Assessment and data-driven modeling of subgrid-scale thermophysical properties in supercritical channel flows

Challenges in large eddy simulation (LES) of wall-bounded turbulent flow and heat transfer at supercritical pressure (SCP) arise from highly nonlinear variations of thermophysical properties. The applicability of LES methodology developed for low-pressure, ideal-gas flows to SCP fluid flows remains uncertain. To address this, the LES framework and assumptions for turbulent channel flows at SCP are carefully evaluated, accompanied by a comprehensive a priori analysis and data-driven modeling. Direct numerical simulations (DNS) are conducted using CO₂ as the working fluid, with different temperature differences between the upper and lower walls. An a priori assessment of subgrid-scale (SGS) terms is performed against the DNS results at different filter widths, through relative error and order-of-magnitude analyses. The findings underscore the importance of SGS thermophysical properties at SCP, with up to 25 % in filtered density, 35 % in filtered dynamic viscosity, and 90 % in filtered thermal diffusivity, giving rise to additional unresolved convective and diffusive fluxes. The feedforward neural networks, with input parameters rigorously selected via Pearson and Spearman correlation analyses, are developed to predict SGS properties. Good performance is achieved, with mean relative errors <10 % for all SGS properties, showing their promising application in SCP CO₂ wall-bounded flows.