Data-driven correction of a turbulence model for separation-induced and natural boundary-layer transition

Accurate prediction of transitional boundary layers remains a persistent challenge for conventional Reynolds-averaged Navier–Stokes turbulence models. This study presents a data-driven correction of a conventional turbulence model for two representative boundary-layer transition scenarios: separation-induced transition and natural transition in attached flow. The Spalart–Allmaras model is used as a representative baseline model to demonstrate the proposed correction framework. Field inversion is first performed to infer a spatially varying correction to the turbulence-production term, with objective functions constructed from available case-specific reference data. The Sandia 809 (S809) airfoil is considered for separation-induced transition, whereas the Natural-Laminar-Flow (First Series)-0416 (NLF(1)-0416) airfoil is used for attached-flow natural transition. The inferred corrections suppress upstream turbulence production, restore upstream laminar boundary-layer regions, and capture transition locations over a range of angles of attack for both airfoils, thereby leading to more accurate predictions of aerodynamic loads, including quantities not directly constrained in the corresponding inversion objectives. A neural network is then trained to represent the correction as a function of local flow features. The trained correction is evaluated on a range of cases not included in training, spanning variations in angle of attack and Reynolds number as well as an unseen airfoil geometry. The transition-aware correction preserves turbulent boundary-layer behavior without producing spurious transition-like responses in nominally turbulent regions. This study demonstrates that data-driven model correction can enable a conventional turbulence model to predict transitional boundary-layer flows in the scenarios considered here, without introducing additional transition-specific transport equations.