Projecte llegit
Títol: Exploration of Machine Learning Data-Driven Models for Aeronautical Applications
Estudiants que han llegit aquest projecte:
GARCÍA PALLARÈS, JUDIT (data lectura: 15-07-2025)- Cerca aquest projecte a Bibliotècnica
GARCÍA PALLARÈS, JUDIT (data lectura: 15-07-2025)- Cerca aquest projecte a Bibliotècnica


Director/a: ROSSI, RICCARDO
Departament: DECA
Títol: Exploration of Machine Learning Data-Driven Models for Aeronautical Applications
Data inici oferta: 29-01-2025 Data finalització oferta: 29-09-2025
Estudis d'assignació del projecte:
DG ENG AERO/SIS TEL
DG ENG AERO/TELEMÀT
DG ENG SISTE/TELEMÀT
Tipus: Individual | |
Lloc de realització: EETAC | |
Nom del segon director/a (UPC): Marco Zuñiga Perez | |
Departament 2n director/a: | |
Paraules clau: | |
Machine Learning, Data Driven, Non-intrusive, Reduced Order Modelling | |
Descripció del contingut i pla d'activitats: | |
The work is dedicated to investigating the potential application, for problems relevant for Aeronautical Engineering, of machine learning data-driven models also known as non-intrusive reduced order models (ROMs). These methods consist in performing the training of a machine learning model using as training data the results of a computationally expensive numerical model, for a set ofrelevant parameters. Once the training has been successful, the trained model is used to efficiently obtain approximated solutions for unseen parameters.
Objectives and Tasks 1. Familiarization with the relevant literature on Machine Learning Data-Driven Models. 2. Development of a comprehensive framework using libraries such as EZyRB, TensorFlow, PyTorch, etc. The framework should incorporate among others, the following methodologies: ' Auto Encoders (AE) ' Radial Basis Functions (RBF) ' Proper Orthogonal Decomposition (POD) ' Gaussian Process Regresion (GPR) 3. Apply the developed framework on a set of benchmark Examples. Evaluate the performance of the framework using metrics for speedup, accuracy and efficiency. 4. Produce a demonstrator using the developed framework and a library for visualization of numerical simulations' results (e.g. PyVista, vedo) |
|
Overview (resum en anglès): | |
This project explores the use of machine learning algorithms to predict the pressure distribution around aerodynamic geometries. A data-driven model will be developed to estimate the pressure coefficient, using the angle of attack and Mach number as inputs. This new model will reduce computational cost by decreasing the dependence on aerodynamic simulations.
The methodology combines several techniques: Radial Basis Function (RBF) interpolation technique using linear and Gaussian basis functions, Proper Orthogonal Decomposition (POD) via Singular Value Decomposition (SVD), and the K-means clustering algorithm to improve interpolation in more complex regimes. The interpolation will be applied to a 2D and 3D domain. For the 2D case, a database of 2000 simulations is generated, simulating pressure distributions around the NACA0012's profile in a two-dimensional space using various angle of attack and Mach number combinations. For the 3D case, the database consists of 500 simulations around the Onera M6 Wing profile in a fully three-dimensional domain. The dataset is split into training, validation, and testing subsets. Results for 2D interpolation show that linear RBF performs well when shock waves are not present, with aerodynamic parameter errors typically below 1%. Gaussian RBF performs worse, especially in transonic regimes, due to numerical instability and lack of regularization. In 2D, applying SVD does not consistently improve results and offers minimal computational gains, making its use optional. In contrast, for 3D interpolation, the use of SVD is needed due to the higher complexity of the model. As in the 2D case, results show that linear RBF performs well in the different studied cases, and that gaussian RBF performs worse, especially when shock waves are present. When K-means is applied, results improve significantly for the gaussian RBF. In summary, this project demonstrates that machine learning techniques can be used to predict the pressure distribution around geometries in both 2D and 3D domains. |