The finite element method is one of the most popular approaches to solve partial differential equations in engineering and physics by discretizing a continuous domain into smaller elements. It offers high accuracy in a wide range of problems, especially in the fields of structural mechanics, fluid mechanics, and heat transfer. However, it can be computationally expensive and lacks flexibility and generalization capabilities. Therefore, alternative methods have been explored and developed. Among them, one of the most innovative and promising approaches is physics-informed neural networks, PINNs, a class of artificial neural networks in the field of machine learning, where the physics of the problem is directly embedded in the network. Embedding physics improves the generalization capability typical of machine learning by improving the information content of the data and finding solutions that intrinsically satisfy the governing physics. This work presents an energy-based physics-informed neural network built on the extreme learning machine paradigm. This approach significantly reduces the computational effort compared to standard artificial neural network training, while the energy-based loss function allows the model to perform well even with very little labeled data, which is particularly relevant in real-world applications where data collection can be challenging. The energy-based framework is validated on a problem involving a Kirchhoff plate and compared to a similar framework based on the strong formulation - PDEs level - of the loss function. The energy-based PINN demonstrates greater stability than the strong-based PINN when the network architecture or the number of available points is changed, but is outperformed when solving a classical laminate theory problem. However, it shows substantial improvement when even a small amount of labeled data is used, making this technology a valid alternative to FEM and worthy of further developments and refinements, some of which are suggested at the end of this work.
Il metodo degli elementi finiti è uno degli approcci più popolari per risolvere equazioni differenziali parziali in ingegneria e fisica discretizzando un dominio continuo in elementi più piccoli. Offre un'elevata accuratezza in un'ampia gamma di problemi, in particolare nei campi della meccanica strutturale, della meccanica dei fluidi e del trasferimento di calore. Tuttavia, può essere computazionalmente costoso e manca di flessibilità e generalizzazione. Pertanto, sono stati sviluppati metodi alternativi. Tra questi, uno degli approcci più innovativi e promettenti è quello delle reti neurali informate dalla fisica, PINN, una classe di reti neurali artificiali nel campo del machine learning, in cui la fisica del problema è incorporata nella rete. Incorporare la fisica migliora la capacità di generalizzazione tipica del machine learning, favorendo l'apprendimento dai dati e trovando soluzioni che soddisfano intrinsecamente le equazioni. Questo lavoro presenta una rete neurale a fisica informata basata sulla formulazione energetica delle equazioni del problema, costruita sul paradigma dell'extreme machine learning. Questo approccio riduce significativamente lo sforzo computazionale rispetto all'addestramento standard della rete neurale, mentre la funzione di perdita basata sull'energia consente al modello di funzionare bene anche con pochissimi dati etichettati, il che è particolarmente rilevante nelle applicazioni in cui la raccolta degli stessi può essere impegnativa. Il framework basato sull'energia è validato su un problema che coinvolge una piastra di Kirchhoff e confrontato con un framework basato sulla forma forte del problema, cioè minimizzando un residuo ricavato dalle equazioni. La PINN basata sull'energia dimostra una maggiore stabilità rispetto alla PINN basata sulla forma forte quando l'architettura della rete o il numero di collocation points vengono modificati, ma viene superata quando si affronta il problema di un laminato in comportamento flessionale. Tuttavia, mostra un miglioramento sostanziale quando viene utilizzata anche una piccola quantità di dati etichettati, rendendo questa tecnologia una valida alternativa al FEM e degna di ulteriori sviluppi, alcuni dei quali suggeriti alla fine di questo lavoro.
An application of extreme learning PINNs to thin plate analysis
Guerriero, Antonio
2023/2024
Abstract
The finite element method is one of the most popular approaches to solve partial differential equations in engineering and physics by discretizing a continuous domain into smaller elements. It offers high accuracy in a wide range of problems, especially in the fields of structural mechanics, fluid mechanics, and heat transfer. However, it can be computationally expensive and lacks flexibility and generalization capabilities. Therefore, alternative methods have been explored and developed. Among them, one of the most innovative and promising approaches is physics-informed neural networks, PINNs, a class of artificial neural networks in the field of machine learning, where the physics of the problem is directly embedded in the network. Embedding physics improves the generalization capability typical of machine learning by improving the information content of the data and finding solutions that intrinsically satisfy the governing physics. This work presents an energy-based physics-informed neural network built on the extreme learning machine paradigm. This approach significantly reduces the computational effort compared to standard artificial neural network training, while the energy-based loss function allows the model to perform well even with very little labeled data, which is particularly relevant in real-world applications where data collection can be challenging. The energy-based framework is validated on a problem involving a Kirchhoff plate and compared to a similar framework based on the strong formulation - PDEs level - of the loss function. The energy-based PINN demonstrates greater stability than the strong-based PINN when the network architecture or the number of available points is changed, but is outperformed when solving a classical laminate theory problem. However, it shows substantial improvement when even a small amount of labeled data is used, making this technology a valid alternative to FEM and worthy of further developments and refinements, some of which are suggested at the end of this work.File | Dimensione | Formato | |
---|---|---|---|
2025_04_Guerriero.pdf
accessibile in internet per tutti
Descrizione: Testo della tesi
Dimensione
2.69 MB
Formato
Adobe PDF
|
2.69 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/10589/234256