In the context of non-invasive hemodynamic assessment, physics informed neural networks (PINNs) represent a highly promising algorithm for enhancing measurement methodologies. They inherently combine the interpolating nature of neural networks with the modelling capabilities of partial differential equations (PDEs). Since their introduction in deep learning field, various applications have emerged in fluid dynamics. This present work aimed to address the lack of systematic studies on the influence of hyperparameters combinations, in relation to the intrinsic characteristics of the application, on PINNs behavior. Through personalized hyperparameter search and automated hyperparameter optimizations, two formulations were investigated concurrently: one with surrogate modelling without data within the domain (NSS) and another with sparse observations (FSSO) in idealized 2D cases. Subsequently, a 3D velocity and pressure field reconstruction problem was tackled using the same formulations in a domain with real geometry. The conducted study revealed the validity of systematic hyperparameter analysis within a specific application, as recognizable and more or less complex patterns of influence were identified. These findings were used to discern relations that were preserved or altered in the 3D context, enabling agile exploration of optimal hyperparameter combinations. In future developments, this work could be further extended to address the time-varying problem and incorporate real velocity data.
Nella valutazione emodinamica non invasiva, le reti neurali fisicamente informate (PINN) rappresentano un algoritmo molto promettente per migliorare le tecniche di misurazione. Esse sono in grado di combinare intrinsecamente la natura interpolatoria delle reti neurali con la modellizzazione fisica espressa delle equazioni differenziali parziali (PDE). A seguito della loro introduzione, sono emerse diverse applicazioni in fluidodinamica. Il presente lavoro è stato condotto per colmare la mancanza di studi sistematici sull’influenza delle combinazioni di iperparametri, in relazione alle caratteristiche intrinseche dell’applicazione, sul comportamento delle PINN. Attraverso una ricerca personalizzata e ottimizzazioni automatizzate di iperparametri, sono state studiate contemporaneamente due formulazioni: una con modellizzazione surrogata senza dati all’interno del dominio (NSS) e un’altra con osservazioni sparse (FSSO), in casi 2D idealizzati. Successivamente, è stato affrontato un problema di ricostruzione del campo di velocità e pressione, in un dominio con geometria reale 3D, impostando le stesse formulazioni. Lo studio condotto ha rivelato la validità dell’analisi sistematica degli iperparametri nell’ambito di un’applicazione specifica, poiché sono stati identificate varie relazioni di influenza. Questi risultati sono stati utilizzati per discernere le relazioni che si sono conservate o alterate nel contesto 3D, consentendo un’esplorazione facilitata delle combinazioni ottimali di iperparametri. In futuro, questo lavoro potrebbe essere ulteriormente ampliato per rendere il problema tempo dipendente e incorporare dati di velocità reali.
On the behavior of physics informed neural networks in hemodynamic applications
Alongi, Filippo
2022/2023
Abstract
In the context of non-invasive hemodynamic assessment, physics informed neural networks (PINNs) represent a highly promising algorithm for enhancing measurement methodologies. They inherently combine the interpolating nature of neural networks with the modelling capabilities of partial differential equations (PDEs). Since their introduction in deep learning field, various applications have emerged in fluid dynamics. This present work aimed to address the lack of systematic studies on the influence of hyperparameters combinations, in relation to the intrinsic characteristics of the application, on PINNs behavior. Through personalized hyperparameter search and automated hyperparameter optimizations, two formulations were investigated concurrently: one with surrogate modelling without data within the domain (NSS) and another with sparse observations (FSSO) in idealized 2D cases. Subsequently, a 3D velocity and pressure field reconstruction problem was tackled using the same formulations in a domain with real geometry. The conducted study revealed the validity of systematic hyperparameter analysis within a specific application, as recognizable and more or less complex patterns of influence were identified. These findings were used to discern relations that were preserved or altered in the 3D context, enabling agile exploration of optimal hyperparameter combinations. In future developments, this work could be further extended to address the time-varying problem and incorporate real velocity data.File | Dimensione | Formato | |
---|---|---|---|
2023_10_Executive_Summary_Alongi.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: 2023_10_Executive_Summary_Alongi
Dimensione
3.73 MB
Formato
Adobe PDF
|
3.73 MB | Adobe PDF | Visualizza/Apri |
2023_10_Tesi_Alongi.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: 2023_10_Tesi_Alongi
Dimensione
24.07 MB
Formato
Adobe PDF
|
24.07 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/10589/210832