Predicting vibrational displacement fields in thin plates is essential for numerous engineering applications, including the acoustic characterization of musical instrument soundboards. Conventional approaches, such as high-fidelity finite element simulations, often involve significant computational costs, while experimental methods may require dense spatial measurements. This thesis proposes a Physics-Informed Neural Network (PINN) approach for reconstructing the displacement field of a thin orthotropic wooden plate subjected to a point load from sparse spatio-temporal observations. The model is validated using COMSOL-generated data from a thin rectangular plate with material properties representative of a violin soundboard. By combining simulated data with the governing Kirchhoff-Love partial differential equation, the proposed approach enables physically consistent interpolation of the structural response. A Sinusoidal Representation Network (SIREN) is employed as the underlying architecture for both the PINN and a purely data-driven baseline, ensuring a fair comparison in terms of model capacity. Results demonstrate that incorporating physical constraints significantly improves reconstruction accuracy, particularly when only limited measurement data are available. The framework is further tested in a time-extrapolation scenario, where observations are restricted to an initial temporal window. The PINN maintains physically consistent predictions beyond the observed interval and successfully captures bending-wave propagation, although prediction accuracy gradually deteriorates as the extrapolation horizon increases. Finally, an inverse analysis highlights the intrinsic non-uniqueness of identifying orthotropic stiffness parameters from displacement measurements alone, revealing fundamental limitations in the estimation of the associated mechanical parameters.
La previsione dei campi di spostamento vibrazionale nelle piastre sottili è essenziale per numerose applicazioni ingegneristiche, inclusa la caratterizzazione acustica delle tavole armoniche degli strumenti musicali. Gli approcci convenzionali, come ad esempio le simulazioni agli elementi finiti, spesso comportano costi computazionali significativi, mentre i metodi sperimentali possono richiedere misurazioni molto dense. Questa tesi propone un approccio basato su Physics-Informed Neural Network (PINN) per la ricostruzione del campo di spostamento di una piastra di legno sottile ortotropa sottoposta a un carico puntiforme a partire da osservazioni spazio-temporali sparse. Il modello viene validato utilizzando dati generati in COMSOL relativi a una piastra rettangolare sottile con proprietà dei materiali rappresentative della tavola armonica di un violino. Combinando i dati simulati con l’equazione di Kirchhoff–Love, l’approccio proposto consente un’interpolazione fisicamente coerente. Lo studio propone una Sinusoidal Representation Network (SIREN) come architettura di base sia per la PINN che per una baseline puramente data-driven, garantendo un confronto equo in termini di capacità del modello. I risultati dimostrano che l’inclusione dei vincoli fisici migliora significativamente l’accuratezza della ricostruzione, in particolare quando sono disponibili solo dati di misura limitati. Il framework viene inoltre testato in uno scenario di extrapolazione temporale, in cui i dati disponibili sono limitati solo a una finestra temporale iniziale. La PINN mantiene previsioni fisicamente coerenti oltre l’intervallo osservato e riesce a catturare la propagazione delle onde flessionali, sebbene l’accuratezza della previsione diminuisca gradualmente all’aumentare dell’orizzonte di extrapolazione. Infine, un’analisi inversa evidenzia la non univocità intrinseca nell’identificazione dei parametri di rigidezza ortotropa a partire dalle sole misure di spostamento, rivelando limitazioni fondamentali della stima dei parametri meccanici associati.
Interpolation and extrapolation of time displacement in thin orthotropic plates using Physics-Informed Neural Networks
Benesso, Francesca
2025/2026
Abstract
Predicting vibrational displacement fields in thin plates is essential for numerous engineering applications, including the acoustic characterization of musical instrument soundboards. Conventional approaches, such as high-fidelity finite element simulations, often involve significant computational costs, while experimental methods may require dense spatial measurements. This thesis proposes a Physics-Informed Neural Network (PINN) approach for reconstructing the displacement field of a thin orthotropic wooden plate subjected to a point load from sparse spatio-temporal observations. The model is validated using COMSOL-generated data from a thin rectangular plate with material properties representative of a violin soundboard. By combining simulated data with the governing Kirchhoff-Love partial differential equation, the proposed approach enables physically consistent interpolation of the structural response. A Sinusoidal Representation Network (SIREN) is employed as the underlying architecture for both the PINN and a purely data-driven baseline, ensuring a fair comparison in terms of model capacity. Results demonstrate that incorporating physical constraints significantly improves reconstruction accuracy, particularly when only limited measurement data are available. The framework is further tested in a time-extrapolation scenario, where observations are restricted to an initial temporal window. The PINN maintains physically consistent predictions beyond the observed interval and successfully captures bending-wave propagation, although prediction accuracy gradually deteriorates as the extrapolation horizon increases. Finally, an inverse analysis highlights the intrinsic non-uniqueness of identifying orthotropic stiffness parameters from displacement measurements alone, revealing fundamental limitations in the estimation of the associated mechanical parameters.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_06_Benesso_Executive_Summary.pdf
non accessibile
Descrizione: Executive Summary of the Thesis
Dimensione
2.25 MB
Formato
Adobe PDF
|
2.25 MB | Adobe PDF | Visualizza/Apri |
|
2026_06_Benesso_Thesis.pdf
non accessibile
Descrizione: Text of the Thesis
Dimensione
8.64 MB
Formato
Adobe PDF
|
8.64 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/10589/261550