Development of A Lightweight Neural Network for Predicting Reaction Rates in RANS Simulations of Flame-Wall Interactions

Abstract Accurate prediction of Flame-Wall Interactions (FWI) is crucial for the design of compact, high-efficiency combustors but remains a significant challenge for Reynolds-Averaged Navier-Stokes (RANS) simulations. This study presents the development and validation of a lightweight data-driven closure for the mean reaction rate, trained on high-fidelity Direct Numerical Simulation (DNS) data of a V-shaped turbulent premixed flame. To ensure numerical stability and physical consistency, the training methodology incorporates a logarithmic transformation of the reaction rate target and a novel error-function transformation of the wall distance, effectively localizing the model’s sensitivity to the near-wall quenching zone. A-priori validation demonstrates that the neural network accurately captures the mean reaction rate statistics and extrapolates successfully to unseen spatial regions. However, a-posteriori assessment in coupled RANS simulations reveals a critical sensitivity to input feature consistency. The full-feature model (7 input parameters), which includes turbulent kinetic energy and dissipation rate as inputs, fails to predict the correct quenching behavior due to significant discrepancies between the DNS-extracted turbulence quantities used for training and the corresponding quantities obtained from RANS. In contrast, a reduced-feature model (3 input parameters), relying exclusively on the mean progress variable, its gradient, and the transformed wall distance, demonstrates remarkable robustness. By decoupling the prediction from uncertain turbulence quantities, the model with 3 input parameters accurately reproduces the V-flame structure, achieving performance comparable to a manually tuned algebraic closure. Furthermore, the reduced model exhibits strong generalization capabilities, successfully predicting the transient physics of a Head-On Quenching (HOQ) configuration that was not present in the training dataset. These findings highlight that for data-driven RANS closures, selecting input features that remain consistent across fidelity levels, is paramount for achieving predictive stability and generalization.