This thesis presents a novel approach for the multiscale analysis of inelastic microstructured materials. The result of the study is a homogenization technique based on the combination of machine learning tools. Dimensionality reduction techniques, neural networks and system dynamics identification methods are exploited and integrated with thermodynamics principles. The primary goal of this research is the development of a method, which can accurately predict the constitutive response of complex materials at the macroscopic scale. The method allows for unsupervisedly select features of a microstructured material model for serving as Internal State Variables at the macroscopic state. Thermodynamics-based artificial neural networks are used to learn the macroscopic constitutive response, in compliance with thermodynamics laws. The response of the investigated systems to external excitation is also mapped by evolution laws. The latter are learned from data with the aid of the dynamic mode decomposition with control, a dynamic system identification method. The latter is modified to achieve the stability of the integration scheme in time and to increase the accuracy. A computational toolbox has been developed to aid the analysis of continuous microsystems modeled as representative unit cells or volumes. The toolbox allows for generating periodic models, periodic mesh, and assigning periodic boundary conditions, taking advantage of the commercial software ABAQUS.
Questa tesi propone un nuovo approccio per l’analisi multiscala di materiali inelstici microstrutturati. Il risultato dello studio è una tecnica di omogeneizzazione basata sulla combinazione di strumenti di Machine Learning, come le tecniche di riduzione della dimensionalità e le reti neurali artificiali, metodi di identificazione della dinamica dei sistemi, come la Dynamic Mode Decomposition e le leggi della termodinamica. L’obiettivo principale della ricerca è lo sviluppo di un metodo in grado di simulare accuratamente la risposta costitutiva di materiali microstrutturati alla scala macroscopica. Il metodo elaborato consente di selezionare in modo non supervisionato le variabili di stato interne da utilizzare alla macroscala. Le reti neurali artificiali basate sulla termodinamica sono utilizzate per apprendere la risposta costitutiva macroscopica in conformità con le leggi della termodinamica. Il comportamento inelastico dei sistemi investigati in risposta alle sollecitazioni esterne è tracciato dalle leggi di evoluzione. Queste ultime vengono apprese dai dati utilizzando la tecnica chiamta Dynamic Mode Decomposition with Control. Questa è stata estesa nel lavoro di tesi affinché restituisca degli operatori in grado di tracciare una dinamica stabile del sistema indagato e affinchè le traiettorie predette siano accurate. L’analisi dei sistemi microscopici considerati è stata effettuata grazie allo sviluppo di un toolbox computazionale per l’analisi di celle unitarie rappresentative periodiche. Il toolbox consente di generare modelli periodici, mesh periodiche e assegnare condizioni al contorno periodiche, all’interno dell’ambiente di calcolo del software commerciale ABAQUS.
Homogenization of microstructured materials via thermodynamics based artificial neural networks and dimensionality reduction techniques
PIUNNO, GIOVANNI
2023/2024
Abstract
This thesis presents a novel approach for the multiscale analysis of inelastic microstructured materials. The result of the study is a homogenization technique based on the combination of machine learning tools. Dimensionality reduction techniques, neural networks and system dynamics identification methods are exploited and integrated with thermodynamics principles. The primary goal of this research is the development of a method, which can accurately predict the constitutive response of complex materials at the macroscopic scale. The method allows for unsupervisedly select features of a microstructured material model for serving as Internal State Variables at the macroscopic state. Thermodynamics-based artificial neural networks are used to learn the macroscopic constitutive response, in compliance with thermodynamics laws. The response of the investigated systems to external excitation is also mapped by evolution laws. The latter are learned from data with the aid of the dynamic mode decomposition with control, a dynamic system identification method. The latter is modified to achieve the stability of the integration scheme in time and to increase the accuracy. A computational toolbox has been developed to aid the analysis of continuous microsystems modeled as representative unit cells or volumes. The toolbox allows for generating periodic models, periodic mesh, and assigning periodic boundary conditions, taking advantage of the commercial software ABAQUS.File | Dimensione | Formato | |
---|---|---|---|
PhD_Thesis_Giovanni_Piunno.pdf
accessibile in internet per tutti
Descrizione: PhD Thesis Giovanni Piunno
Dimensione
11.11 MB
Formato
Adobe PDF
|
11.11 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/10589/216295