Ofertes treballs fi d'estudishttps://eseiaat.upc.edu/ca/curs-actual/treballs-fi-estudis/ofertes-treballs-fi-destudishttps://eseiaat.upc.edu/++resource++plone-logo.svg
2010 - GRAU EN ENGINYERIA EN TECNOLOGIES AEROESPACIALS
2010 - GRAU EN ENGINYERIA EN VEHICLES AEROESPACIALS
Identificador de l'oferta:
205-06045
Modalitat:
Universitat
Possibilitat de beca/finançament:
No
Idioma d'elaboració del treball:
Anglès
Descripció:
Large-scale structural analysis often requires iterative methods (e.g., Conjugate Gradient, GMRES, BiCGSTAB) to solve high-dimensional systems of equations. These solvers can be very time-consuming, especially in nonlinear or complex structural problems. This project investigates how machine learning (ML) techniques—such as neural networks, supervised methods, or reinforcement learning—can accelerate solver convergence by dynamically optimizing parameters, preconditioners, or convergence criteria. The ultimate goal is to reduce computational cost while maintaining solution accuracy and reliability.
Objectius:
Literature Review:
Survey state-of-the-art iterative solvers for structural analysis and recent ML applications in simulation and numerical methods.
Key Parameter Identification:
Determine which solver parameters and structural features (e.g., stiffness matrix properties, boundary conditions, mesh density) most influence convergence.
ML Model Development:
Design and train an ML model (e.g., neural network, gradient boosting, reinforcement learning) to predict or adjust solver strategies.
Implementation and Integration:
Incorporate the ML component into an existing solver framework or a custom computational code, ensuring compatibility and efficiency.
Validation:
Test and compare the enhanced solver against standard approaches, evaluating convergence speed, accuracy, and computational cost.
Documentation:
Present the methodology, experimental setup, results, and conclusions, highlighting limitations and potential improvements.
Tasques a realitzar / Característiques:
Preliminary Research:
Study relevant literature on iterative solvers and ML-based solver acceleration.
Understand existing tools and libraries for numerical methods and ML.
Data Preparation & Experiment Design:
Generate or gather datasets representing various structural problems.
Define performance metrics (e.g., iteration count, runtime, error tolerance).
Model Training & Testing:
Select and implement an appropriate ML approach.
Tune hyperparameters and validate predictive accuracy on training datasets.
Solver Integration:
Embed the trained model into the solver loop to adapt parameters in real time.
Ensure robust and efficient communication between the solver and ML model.
Analysis & Comparison:
Run simulations on multiple test cases (linear/nonlinear, different scales).
Compare performance with conventional solvers in terms of convergence rate, accuracy, and computational expense.
Final Reporting:
Document all findings, experiments, and results.
Prepare a final report and presentation to showcase the methodology, outcomes, and future directions.
Comparteix: