Accurate numerical simulations can be often computationally expensive and very time consuming. This might become prohibitive whenever we are interested to solve parametrized PDEs several times for different parameter instances as in uncertainty quantification or PDEconstrained optimization problems. On the other hand, less accurate but cheaper models are often available and can provide a large amount of data in a computationally efficient manner. Multifidelity methods combine data with different fidelity levels to deal with such problems. By leveraging the correlations between high- and low-fidelity data sets, multifidelity models manage to achieve high accuracy at a reasonable cost. In the present work, we introduce some of the most recent and efficient artificial neural network architectures for multifidelity regression. Moreover, we compare several possible choices of lowand high-fidelity models (e.g. reduced order models built through the reduced basis method and finite elements models, respectively) and we assess the impact of both the quality and the amount of the training data on the overall accuracy of the numerical outcome obtained through multifidelity neural network regression. Then, we apply this approach to parametrized PDE problems, such as the propagation of a pressure wave into an acoustic horn with parametrized shape and frequency, and a nonlinear elasticity problem in solid mechanics. Finally we extend the multifidelity techniques to time-dependent problems, including a simple nonlinear problem in cardiac electrophysiology, describing the propagation of the action-potential in the cardiac cells.
Simulazioni numeriche accurate spesso possono essere computazionalmente costose e richiedere molto tempo. Ciò potrebbe diventare proibitivo ogni qualvolta siamo interessati a risolvere più volte EDP parametrizzate per diverse istanze dei parametri, come nell’ambito della quantificazione dell’incertezza o nei problemi di controllo ottimo governati da EDP. D’altra parte, sono spesso accessibili modelli meno accurati ma anche meno costosi, che possono fornire una grande quantità di dati in modo più efficace dal punto di vista computazionale. I metodi multifidelity combinano dati a diversi livelli di fedeltà per far fronte a tali problematiche. Sfruttando le correlazioni tra i dati ricavati da modelli ad alta e bassa fedeltà, i modelli multifidelity riescono a raggiungere un’elevata accuratezza con costi ragionevoli. Nel presente lavoro, introduciamo alcune delle più recenti ed efficienti architetture di reti neurali artificiali per affrontare problemi di regressione multifidelity. Inoltre confrontiamo diverse possibili scelte di modelli a bassa e alta fedeltà (ad esempio modelli a basi ridotte e modelli ad elementi finiti, rispettivamente) e valutiamo l’impatto sia della qualità che della quantità dei dati di training sull’accuratezza complessiva del risultato numerico ottenuto attraverso la regressione multifidelity basata su reti neurali. Quindi, applichiamo questo approccio a problemi che coinvolgono EDP parametrizzate, come la propagazione di un’onda di pressione in un corno acustico con forma e frequenza parametrizzate, o un problema di elasticità non lineare nell’ambito della meccanica dei solidi. Infine, estendiamo le tecniche multifidelity a problemi tempo-dipendenti, tra cui un semplice problema nonlineare di interesse in elettrofisiologia cardiaca riguardante la propagazione del potenziale d’azione nelle cellule cardiache.
Multifidelity regression with artificial neural networks : efficient approximation of output quantities for parametrized systems
CONTI, PAOLO
2019/2020
Abstract
Accurate numerical simulations can be often computationally expensive and very time consuming. This might become prohibitive whenever we are interested to solve parametrized PDEs several times for different parameter instances as in uncertainty quantification or PDEconstrained optimization problems. On the other hand, less accurate but cheaper models are often available and can provide a large amount of data in a computationally efficient manner. Multifidelity methods combine data with different fidelity levels to deal with such problems. By leveraging the correlations between high- and low-fidelity data sets, multifidelity models manage to achieve high accuracy at a reasonable cost. In the present work, we introduce some of the most recent and efficient artificial neural network architectures for multifidelity regression. Moreover, we compare several possible choices of lowand high-fidelity models (e.g. reduced order models built through the reduced basis method and finite elements models, respectively) and we assess the impact of both the quality and the amount of the training data on the overall accuracy of the numerical outcome obtained through multifidelity neural network regression. Then, we apply this approach to parametrized PDE problems, such as the propagation of a pressure wave into an acoustic horn with parametrized shape and frequency, and a nonlinear elasticity problem in solid mechanics. Finally we extend the multifidelity techniques to time-dependent problems, including a simple nonlinear problem in cardiac electrophysiology, describing the propagation of the action-potential in the cardiac cells.File | Dimensione | Formato | |
---|---|---|---|
2021_4_Conti.pdf
solo utenti autorizzati dal 07/04/2022
Descrizione: Testo della tesi
Dimensione
4.85 MB
Formato
Adobe PDF
|
4.85 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/10589/173600