DEVELOPMENT OF SURROGATE MODEL USING DEEP LEARNING AND PROPER ORTHOGONAL DECOMPOSITION METHOD ON AEROELASTIC DATA

• Main Author: Iksan Musyahada, Kukuh
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/70223
• Institution: Institut Teknologi Bandung

Where non-linear aerodynamics sets a precedent in the governing flow, to which the transonic mach regime is concerned, producing aeroelastic models under these conditions is very expensive. This thesis investigates the utilization of Deep Learning to produce a surrogate model response surface based on the single fidelity aeroelastic simulation result to predict the dynamic response of an airfoil. Moreover, the aeroelastic analysis is limited to predicting dynamic response such as Lift Coefficient (Cl), Drag Coefficient (Cd), and the airfoil motion (Plunge and Pitch) within the subsonic and transonic mach regime based on a NACA 64A010 Airfoil. This thesis is chosen as an effort to minimize computational cost. Whereby, the proposed space is bounded and varied in the Mach range of 0.6-0.8 and the Flutter Speed Index of a range of 0.4-2.0. This work will utilize Proper Orthogonal Decomposition (POD) to process the data. The POD method will process the data by decomposing the aeroelastic data into a set of coefficients and use it as a label for the deep learning. The deep learning model consists of a multi-layer feed-forward neural network that maps each mach number and flutter speed index with the POD coefficients and predicts new cases. The result showed great similarity with simulation data with MAE number less than 10e-2 and MAPE around 50%. From the result the flutter boundary also being generated. The flutter boundary from model proven capable to follow the previous research. Moreover, running the model only needed a maximum of 2 minutes so it will save until 100 times the simulation time.