Perpustakaan Digital ITB

Turbulence is a physical phenomenon that seems straightforward, but remains deeply complex and unclear. Reynolds-Averaged Navier-Stokes (RANS) remain the standard tool for turbulence prediction, yet it has predictive limitations due to modelling constraints. This thesis develops deep learning framework to enhance the ? model for flow over periodic hills. Baseline simulation is conducted in OpenFOAM with selected geometries, from which the flow features are extracted to train deep learning models. The deep learning results are then evaluated on unseen geometries, with the predicted Reynolds stresses subsequently deployed back to the solver. This investigates the utility of deep learning models as comprehensive turbulence model correction, rather than data-driven predictions only. The two implemented deep learning frameworks are Feedforward Neural Network and Kolmogorov-Arnold Network. KAN outperforms FFNN by reaching a normalized-root-mean-squared-error of 0.6781 against 0.7527, while using only two hidden layers. The model enhancement is concentrated in behind the hill region, where the flow separates and recirculates and the baseline model is least reliable, although the two networks correct parts of the flow differently. The results show that deep learning can enhance an established turbulence model on geometries it has not been trained on, with KAN achieving it in a smaller network.