Ofertes treballs fi d'estudis

Títol: Study of Machine-Learning-Based Nonlinear Techniques for Acceleration of Physical Simulations

Director/a: Hernandez Ortega, Joaquin

Email del professor/a: joaquin.alberto.hernandez@upc.edu

Departament del professor/a: 737-RMEE

• 2010 - GRAU EN ENGINYERIA EN TECNOLOGIES AEROESPACIALS
• 2010 - GRAU EN ENGINYERIA EN VEHICLES AEROESPACIALS
• 2014 - MÀSTER UNIVERSITARI EN ENGINYERIA AERONÀUTICA
• 2016 - MÀSTER UNIVERSITARI EN ENGINYERIA ESPACIAL I AERONÀUTICA

Identificador de l'oferta: 205-06404

Modalitat: Universitat

Possibilitat de beca/finançament: No

Idioma d'elaboració del treball: Anglès

Paraules clau:

Descripció:

High-fidelity numerical simulations are essential in many engineering fields, including aeronautics, where accurate predictions of structural deformation, unsteady aerodynamics, thermal behaviour, or fluid–structure interaction often require models with millions of degrees of freedom. Although these simulations provide excellent accuracy, they are computationally expensive and difficult to deploy in contexts requiring fast response, parametric exploration, control, or real-time digital twins. Reduced-Order Models (ROMs) provide an efficient alternative by replacing the full field of nodal values with a small set of reduced coordinates obtained through techniques such as Proper Orthogonal Decomposition (POD). However, for strongly nonlinear regimes—common in aeronautics, e.g. convection-dominated flows, vortex shedding, or nonlinear structural dynamics—classical linear ROMs struggle to remain accurate. A promising direction is to introduce nonlinear relationships between reduced coordinates and to use machine learning tools (neural networks, radial basis functions, etc.) to model these nonlinear dependencies. This project will explore such nonlinear techniques to improve accuracy and accelerate physical simulations relevant to aeronautical engineering, though the approach remains general and broadly applicable. The project can be implemented in Python or MATLAB.

Objectius:

Understand POD-based reduced-order models and their limitations in nonlinear regimes. Implement projection-based ROMs (Galerkin / LSPG) for simple physical problems. Introduce nonlinear dependencies between reduced coordinates using machine learning tools. Evaluate improvements in accuracy and computational efficiency. Optionally apply the methodology to aeronautical-related cases (e.g., convection problems, structural dynamics, simple aerodynamic surrogates). Produce a well-structured, reproducible numerical study.

Tasques a realitzar / Característiques:

Literature review: POD, projection ROMs, closure models, machine learning techniques. Implementation of a full-order reference solver (provided or simplified). Construction and analysis of POD bases and reduced coordinates. Development of nonlinear ROMs using neural networks or other ML regression methods. Numerical experiments on benchmark problems and, optionally, aerospace-inspired cases. Systematic comparison: accuracy, robustness, and speedup. Final report and oral presentation.

Tema: ENGINYERIA AERONÀUTICA

Tipus: Estudi

Càrrega de treball:

La càrrega de treball s'adaptarà als crèdits de la titulació.

Requisits:

Basic knowledge of numerical methods, linear algebra and differential equations. Some previous exposure to the finite element method (FEM) is recommended, at least at an introductory level (assembly concepts, shape functions, discretisation of PDEs). Programming experience in Python or MATLAB is desirable. Interest in computational modelling, physical simulation, or machine learning is a plus.

Temàtica ambiental: No

Confidencial (informatiu): No

Ambit de cooperació: No

Publicació: 09/01/2026

Caducitat: 03/12/2026