The analysis of multilayered composite structures presents an ever increasing challenge with the appearance of new levels of complexity within material science. Proper tools are in constant development to give researchers the capabilities to better understand the nature of materials and how to use them. A novel methodology in the field of Machine Learning (ML), known as Physics Informed Neural Networks (PINN) has been implemented. This architecture is employed to approximate the solution of boundary-value collocation problems formulated in compliance with the Sublaminate Generalized Unified Formulation (S-GUF). The novelty of the approximation resides in the joint application of PINN and the Extreme Learning Machine (ELM) algorithm, which is based on the solution of a minimum-norm least squares linear system to fulfill the physical constraints imposed by the problem. An additional layer of novelty makes it the first attempt to use these techniques in a combined fashion to a variable-kinematics formulation, without previous work reported in the area. The aim of this approach is to gain access to a new tool that allows the analysis of complex multilayered plates with different variable kinematics models in an efficient way, allowing the comparison of methods and models. A special attention is given to the comparison and stability of strong and weak-form formulations within this framework. A strong approach is developed by extracting a system of partial differential equations from the expression of the Principle of Virtual Works, that are imposed across all points in a collocation grid to enforce equilibrium and compatibility conditions. Certain numerical issues are reported along the work, motivating the development of new ways to overcome them. The coupled influence of network parameters and other factors given by the geometry or the approximation is studied thanks to a detailed sensitivity analysis. Finally, the initial implementation made for rectangular plates is extended to shells, and plates of arbitrary shapes, what makes it a versatile tool that overcomes the limitations of other methodologies only suited to solve simple problems.
L'analisi delle strutture composite multistrato rappresenta una sfida sempre più grande con la comparsa di nuovi livelli di complessità nella scienza dei materiali. Strumenti adeguati sono in costante sviluppo per dare ai ricercatori le capacità di comprendere meglio la natura dei materiali e come utilizzarli. È stata implementata una nuova metodologia nel campo del Machine Learning (ML), conosciuta come Physics Informed Neural Networks (PINN). Questa architettura è impiegata per approssimare la soluzione di problemi di collocazione dei valori limite formulati in conformità con la Formulazione Unificata Generalizzata del Sublaminato (S-GUF). La novità dell'approssimazione risiede nell'applicazione congiunta di PINN e dell'algoritmo Extreme Learning Machine (ELM), che si basa sulla soluzione di un sistema lineare dei minimi quadrati a norma minima per soddisfare i vincoli fisici imposti dal problema. Un ulteriore livello di novità lo rende il primo tentativo di utilizzare queste tecniche in modo combinato a una formulazione di cinematica variabile, senza precedenti lavori riportati nell'area. Lo scopo di questo approccio è quello di ottenere l'accesso a un nuovo strumento che permetta l'analisi di piastre complesse multistrato con diversi modelli di cinematica variabile in modo efficiente, permettendo il confronto di metodi e modelli. Un'attenzione speciale è data al confronto e alla stabilità delle formulazioni forti e deboli in questo quadro. Un approccio forte è sviluppato estraendo un sistema di equazioni differenziali parziali dall'espressione del Principio dei Lavori Virtuali, che sono imposte attraverso tutti i punti in una griglia di collocazione per imporre condizioni di equilibrio e compatibilità. Alcuni problemi numerici sono riportati lungo il lavoro, motivando lo sviluppo di nuovi modi per superarli. L'influenza accoppiata dei parametri di rete e di altri fattori dati dalla geometria o dall'approssimazione è studiata grazie a una dettagliata analisi di sensibilità. Infine, l'implementazione iniziale fatta per piastre rettangolari è estesa a gusci e piastre di forme arbitrarie, il che lo rende uno strumento versatile che supera i limiti di altre metodologie adatte solo a risolvere problemi semplici.
Analysis of composite laminates using physics informed neural networks and extreme learning machine
BARCELONA MORENO, IVÁN
2020/2021
Abstract
The analysis of multilayered composite structures presents an ever increasing challenge with the appearance of new levels of complexity within material science. Proper tools are in constant development to give researchers the capabilities to better understand the nature of materials and how to use them. A novel methodology in the field of Machine Learning (ML), known as Physics Informed Neural Networks (PINN) has been implemented. This architecture is employed to approximate the solution of boundary-value collocation problems formulated in compliance with the Sublaminate Generalized Unified Formulation (S-GUF). The novelty of the approximation resides in the joint application of PINN and the Extreme Learning Machine (ELM) algorithm, which is based on the solution of a minimum-norm least squares linear system to fulfill the physical constraints imposed by the problem. An additional layer of novelty makes it the first attempt to use these techniques in a combined fashion to a variable-kinematics formulation, without previous work reported in the area. The aim of this approach is to gain access to a new tool that allows the analysis of complex multilayered plates with different variable kinematics models in an efficient way, allowing the comparison of methods and models. A special attention is given to the comparison and stability of strong and weak-form formulations within this framework. A strong approach is developed by extracting a system of partial differential equations from the expression of the Principle of Virtual Works, that are imposed across all points in a collocation grid to enforce equilibrium and compatibility conditions. Certain numerical issues are reported along the work, motivating the development of new ways to overcome them. The coupled influence of network parameters and other factors given by the geometry or the approximation is studied thanks to a detailed sensitivity analysis. Finally, the initial implementation made for rectangular plates is extended to shells, and plates of arbitrary shapes, what makes it a versatile tool that overcomes the limitations of other methodologies only suited to solve simple problems.File | Dimensione | Formato | |
---|---|---|---|
2021_07_Barcelona.pdf
accessibile in internet per tutti
Descrizione: Barcelona_MSc_thesis_2021
Dimensione
9.93 MB
Formato
Adobe PDF
|
9.93 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/10589/177762