In numerous engineering applications, the computational cost associated with solving physical problems using Full Order Models (FOMs) is impractical, especially in the initial design phase when multiple simulations are required. Surrogate models are developed to alleviate the computational burden, representing a trade-off between accuracy and efficiency. Neural operators, a subset of Deep Learning models, have demonstrated significant potential in constructing surrogates for tackling complex physical problems. In this study, the performance of Feedforward Neural Networks (FNN), Deep Operator Networks (DeepONet), and Fourier Neural Operator (FNO) is evaluated in predicting the material response of a graphite wedge subject to varying boundary conditions applied on its tip. The comparison of results obtained with the different neural operators is utilized to draw theoretical observations. In the examined problem, the DeepONet architecture exhibits superior accuracy and speed-up capabilities. Among the presented neural operators, DeepONet excels in providing a low-rank approximation of this specific case study, enabling the use of a limited number of parameters with respect to other surrogates, while achieving high accuracy. The surrogate achieves a computational time of up to 10,000 times faster than the traditional solver.
In numerose applicazioni ingegneristiche, il costo computazionale associato alla risoluzione di problemi fisici utilizzando Full Order Models (FOM) è limitante, specialmente nelle fasi iniziali di design, in cui sono necessarie svariate simulazioni. I modelli surrogati vengono sviluppati per ridurre l’impatto del costo computazionale, essendo un compromesso tra l’accuratezza e l’efficienza computazionale. I neural operators, una sotto classe dei modelli di Deep Learning, hanno dimostrato ottime potenzialità nella costruzione di surrogati per la modellazione di problemi fisici complessi. In questo studio vengono comparate le performance di Feedforward Neural Networks (FNN), Deep Operator Networks (DeepONet) e Fourier Neural Operator (FNO) per prevedere la risposta termica di un cuneo di grafite, sulla punta del quale vengono applicate diverse condizioni al contorno. La comparazione dei risultati numerici ottenuti usando differenti neural operators viene utilizzata per evidenziare aspetti teorici. Per il problema investigato, l'architettura di DeepONet mostra un'accuratezza e uno speed-up più elevati. Tra i neural operators presentati, DeepONet mostra le migliori capacità nell’effettuare una compressione di basso rango per lo specifico problema in esame, permettendo di utilizzare un numero limitato di parametri rispetto ad altri surrogati, ottenendo contemporaneamente un’elevata accuratezza. Il surrogato raggiunge uno speed-up computazionale di 10,000 volte superiore rispetto al solutore tradizionale.
A preliminary study of Neural Operator-based surrogates for heat transfer assessment in hypersonic conditions
MIGLIAVACCA, MARCO
2022/2023
Abstract
In numerous engineering applications, the computational cost associated with solving physical problems using Full Order Models (FOMs) is impractical, especially in the initial design phase when multiple simulations are required. Surrogate models are developed to alleviate the computational burden, representing a trade-off between accuracy and efficiency. Neural operators, a subset of Deep Learning models, have demonstrated significant potential in constructing surrogates for tackling complex physical problems. In this study, the performance of Feedforward Neural Networks (FNN), Deep Operator Networks (DeepONet), and Fourier Neural Operator (FNO) is evaluated in predicting the material response of a graphite wedge subject to varying boundary conditions applied on its tip. The comparison of results obtained with the different neural operators is utilized to draw theoretical observations. In the examined problem, the DeepONet architecture exhibits superior accuracy and speed-up capabilities. Among the presented neural operators, DeepONet excels in providing a low-rank approximation of this specific case study, enabling the use of a limited number of parameters with respect to other surrogates, while achieving high accuracy. The surrogate achieves a computational time of up to 10,000 times faster than the traditional solver.File | Dimensione | Formato | |
---|---|---|---|
2024_04_Migliavacca_tesi.pdf
accessibile in internet per tutti
Descrizione: Tesi - formato articolo
Dimensione
9.48 MB
Formato
Adobe PDF
|
9.48 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/10589/218642