Flow and heat transfer predictions of supercritical pressure hydrocarbon fuel in three-dimensional Kagome lattice structures using physics-informed neural network

Abstract

Classical computational fluid dynamics faces prohibitive computational costs due to complex cooling structures and variable operating conditions. This study developed a physics-informed neural network (PINN) driven by Reynolds-averaged Navier–Stokes data to predict the flow and heat transfer of supercritical hydrocarbon fuel in three-dimensional Kagome lattice structures. By enforcing governing physical laws, the PINN effectively suppressed the numerical over-smoothing inherent in purely data-driven artificial neural networks (ANN), precisely preserving large velocity gradients within complex wake recirculation and boundary layer separation regions. Under local parametric extension conditions slightly outside the training bounds, the incorporated physical priors prevented non-physical numerical drift, significantly reducing the temperature prediction error from 2.18% (ANN) to 1.04% (PINN). Across the entire fluid domain, the average relative errors for temperature, velocity magnitude, and pressure predictions were strictly bounded to 2.17%, 2.58%, and 0.016%, respectively. Furthermore, utilizing the trained PINN as a rapid surrogate, this study explicitly quantified the diminishing cooling benefits with increased mass flux. Specifically, an equivalent increase in the coolant supply yields a substantial maximum wall temperature reduction, decreasing from 39.5 to 11 K under low flow rate conditions, whereas this temperature drop becomes gradually attenuated under high flow rate conditions. This framework captures how the convective heat transfer coefficient growth rate follows a decaying trend with increasing mass flux. Ultimately, this physics-driven approach establishes reliable quantitative baselines for performance evaluation, offering guidance for the prediction of flow and thermal fields under applied thermal boundary conditions.