In this thesis, we develop a non-intrusive reduced basis (RB) method for parametrized time-indepedent partial differential equations (PDEs). The proposed method extracts a reduced basis from a collection of high-fidelity solutions via proper orthogonal decomposition (POD) and employs artificial neural networks (ANNs), particularly multi-layer perceptrons (MLPs), to accurately approximate the coefficients of the reduced model. The search for the optimal number of inner neurons and the minimum amount of training samples which avoids overfitting is carried out in the offline phase through an automatic routine, relying upon a joint use of the Latin Hypercube Sampling (LHS) and the Levenberg-Marquardt training algorithm. This guarantees a complete offline-online decoupling, leading to an efficient RB method - referred to as POD-NN - suitable also for nonlinear problems featuring a non-affine parametric dependence. Numerical studies are presented for the linear and nonlinear Poisson equation and for driven cavity viscous flows, modeled through the steady uncompressible Navier-Stokes equations. Both physical and geometrical parametrizations are considered. Several results confirm the accuracy of the POD-NN method and show the substantial speed-up enabled at the online stage with respect to a traditional RB strategy based on the Galerkin projection process.
In questa tesi, si propone un metodo a basi ridotte (reduced basis, RB in inglese) non intrusivo per equazioni alle derivate parziali stazionarie parametriche. Il metodo estrae una base ridotta da un insieme di soluzioni altamente accurate tramite decomposizione ortogonale propria (proper orthogonal decomposition, POD) ed utilizza una rete neurale artificiale (artificial neural network, ANN), in particolare un percettrone multistrato (multi-layer perceptron, MLP), per approssimare accuratamente i coefficienti del modello ridotto. La ricerca del numero ottimale di neuroni per strato nascosto e del numero minimo di esempi di apprendimento necessario per non incorrere in overfitting, viene portata a termine nella fase offline tramite una routine automatica, combinando un campionamento a ipercubi latini (Latin Hypercube Sampling, LHS) con l'algoritmo di apprendimento di Levenberg-Marquardt. Questo garantisce un totale disaccoppiamento tra la fase offline e quella online che porta ad un metodo a basi ridotte efficiente - denominato POD-NN - adatto anche per problemi non lineari con una dipendenza non affine dai parametri. Si presentano studi numerici per l'equazione di Poisson (sia lineare che non lineare) e per fluidi viscosi in cavità, modellati tramite le equazioni di Navier-Stokes stazionarie ed incomprimibili. Diversi risultati confermano l'accuratezza del metodo POD-NN e mostrano il sostanziale vantaggio computazionale offerto nella fase online rispetto ad un metodo a basi ridotte tradizionale, basato su un processo di proiezione alla Galerkin.
Reduced order modeling of nonlinear problems using neural networks
UBBIALI, STEFANO
2016/2017
Abstract
In this thesis, we develop a non-intrusive reduced basis (RB) method for parametrized time-indepedent partial differential equations (PDEs). The proposed method extracts a reduced basis from a collection of high-fidelity solutions via proper orthogonal decomposition (POD) and employs artificial neural networks (ANNs), particularly multi-layer perceptrons (MLPs), to accurately approximate the coefficients of the reduced model. The search for the optimal number of inner neurons and the minimum amount of training samples which avoids overfitting is carried out in the offline phase through an automatic routine, relying upon a joint use of the Latin Hypercube Sampling (LHS) and the Levenberg-Marquardt training algorithm. This guarantees a complete offline-online decoupling, leading to an efficient RB method - referred to as POD-NN - suitable also for nonlinear problems featuring a non-affine parametric dependence. Numerical studies are presented for the linear and nonlinear Poisson equation and for driven cavity viscous flows, modeled through the steady uncompressible Navier-Stokes equations. Both physical and geometrical parametrizations are considered. Several results confirm the accuracy of the POD-NN method and show the substantial speed-up enabled at the online stage with respect to a traditional RB strategy based on the Galerkin projection process.File | Dimensione | Formato | |
---|---|---|---|
2017_07_Ubbiali.pdf
accessibile in internet per tutti
Descrizione: Thesis text
Dimensione
5.21 MB
Formato
Adobe PDF
|
5.21 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/134925