In the field of lightweight structures, advancements in manufacturing have allowed the exploration of variable stiffness panels. To achieve load-bearing structural panels with variable stiffness, an approach is to utilise curvilinear stiffeners, which offer greater design flexibility compared to traditional straight stiffeners, but their effective use necessitates an optimised design. This thesis introduces a novel method for optimising the design of curvilinearly stiffened panels, significantly enhancing computational efficiency compared to existing literature. It uniquely combines a semi-analytical approach based on the Ritz method to evaluate the structural response of the panels and a machine learning surrogate model technique. The primary objective is to outperform traditional approaches to optimise the layout of curvilinear stiffeners in terms of computational speed. Central to this method is the surrogate model, which through its particular structure, leverages a semi-supervised learning strategy which reduces the size of the needed dataset and, thus the overall time required to build the surrogate model. Images which describe the layout of the curvilinear stiffeners are given in input to the surrogate model, its first section is an encoder which, through convolutional neural networks detects the layout features, and then the second section assigns to layout features the corresponding structural response. The encoding portion is trained using image data for which no structural response is needed, thus this strategy utilises the same encoding section to construct multiple surrogate models to predict different structural responses for different boundary and loading conditions. A limited but effective set of labelled data is sufficient to train the remainder of the surrogate model, which is obtained through the semi-analytical method. The results demonstrate a substantial reduction in data acquisition time compared to conventional FEA software-based methods, thus vastly reducing the overall optimisation time. These enhancements could facilitate the exploration of more complex panel designs and larger datasets.
I recenti progressi nel campo delle tecniche di fabbricazione di strutture leggere hanno permesso l’esplorazione di pannelli a rigidità variabile. Un metodo adottato per realizzare pannelli strutturali portanti, ampiamente adottati nell’aerospaziale, con rigidità variabile è utilizzare correnti curvilinei. Sebbene i correnti curvilinei offrano una maggiore flessibilità di progettazione rispetto ai correnti dritti tradizionali, il loro uso efficace richiede una progettazione ottimizzata. Questa tesi introduce un metodo innovativo per ottimizzare la progettazione di pannelli con correnti curvilinei, migliorando notevolmente l’efficienza computazionale rispetto alla letteratura esistente. Il metodo presentato combina per la prima volta un approccio semi-analitico basato sul metodo di Ritz per valutare la risposta strutturale dei pannelli e una tecnica di modello surrogate basato sull’intelligenza artificiale. L’obiettivo principale è sorpassare i metodi tradizionali per ottimizzare la configurazione dei correnti curvilinei in termini di velocità computazionale. Il modello surrogate, elemento principale dell’ottimizzazione, sfrutta un set limitato ma efficace di dati ottenuti attraverso il metodo semi-analitico grazie all’adozione di una strategia di training semi-supervisionata. Questo approccio permette una riduzione in termini di tempo rispetto all’analisi agli elementi finiti. I risultati dimostrano una notevole riduzione del tempo di acquisizione dei dati rispetto ai metodi convenzionali basati su software FEA, riducendo così notevolmente il tempo complessivo di ottimizzazione. Inoltre, la prima parte del modello surrogate è un encoder, allenato per riconoscere le caratteristiche dei correnti curvilinei utilizzando un dataset di sole immagini. L’encoder, una volta allenato può essere utilizzato per costruire diversi modelli surrogati, ognuno con diverse condizioni di carico e di vincolo. Questi miglioramenti potrebbero facilitare l’esplorazione di configurazioni più complesse e l’utilizzo di dataset più vasti.
Optimisation of curvilinearly stiffened panels via deep learning and semi-analytical methods
Monti, Maria Edwina
2023/2024
Abstract
In the field of lightweight structures, advancements in manufacturing have allowed the exploration of variable stiffness panels. To achieve load-bearing structural panels with variable stiffness, an approach is to utilise curvilinear stiffeners, which offer greater design flexibility compared to traditional straight stiffeners, but their effective use necessitates an optimised design. This thesis introduces a novel method for optimising the design of curvilinearly stiffened panels, significantly enhancing computational efficiency compared to existing literature. It uniquely combines a semi-analytical approach based on the Ritz method to evaluate the structural response of the panels and a machine learning surrogate model technique. The primary objective is to outperform traditional approaches to optimise the layout of curvilinear stiffeners in terms of computational speed. Central to this method is the surrogate model, which through its particular structure, leverages a semi-supervised learning strategy which reduces the size of the needed dataset and, thus the overall time required to build the surrogate model. Images which describe the layout of the curvilinear stiffeners are given in input to the surrogate model, its first section is an encoder which, through convolutional neural networks detects the layout features, and then the second section assigns to layout features the corresponding structural response. The encoding portion is trained using image data for which no structural response is needed, thus this strategy utilises the same encoding section to construct multiple surrogate models to predict different structural responses for different boundary and loading conditions. A limited but effective set of labelled data is sufficient to train the remainder of the surrogate model, which is obtained through the semi-analytical method. The results demonstrate a substantial reduction in data acquisition time compared to conventional FEA software-based methods, thus vastly reducing the overall optimisation time. These enhancements could facilitate the exploration of more complex panel designs and larger datasets.File | Dimensione | Formato | |
---|---|---|---|
2023_12_Monti.pdf
Open Access dal 29/11/2024
Descrizione: Tesi Magistrale
Dimensione
4.08 MB
Formato
Adobe PDF
|
4.08 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/10589/214163