Investigation of Neural Network-Based Calibration of the Geko Turbulence Model for High-Lift Airfoil Flow Prediction

Abstract

Accurate prediction of aerodynamic forces for high-lift configurations remains challenging in CFD due to turbulence modeling limitations. This study applies neural network-based calibration to the Generalized κ–ω (GEKO) model to improve flow prediction over a multi-element airfoil. Baseline SST-κ–ω and GEKO models matched experimental lift reasonably well but significantly overpredicted drag at higher angles of attack. Two optimized models were developed: GEKO-NN1 (tuning separation coefficient CSEP) and GEKO-NN2 (optimizing both CSEP and near-wall coefficient CNW). Trained using experimental data at an angle of attack of 8.109°, both models were tested across multiple conditions. GEKO-NN2 delivered the best performance, reducing drag error to 6.45% and consistently improving accuracy at higher angles. It also enhanced pressure coefficient predictions, indicating better flow representation. Overall, neural network-based tuning shows strong potential for improving CFD accuracy and reducing manual calibration in complex aerodynamic flows.