Machine Learning (ML) and Neural Networks (NN) specifically represent one of the most active areas of research in the last year. Computational physics and materials science have not been immune to this: Neural Networks and other ML approaches are being adopted as new tools, which promise to solve problems which have been intractable so far. In this Thesis, some advancements in this respect will be shown, focusing on the possibility of tackling traditional problems from new directions. In particular, the discussion will revolve around how to adapt pure Deep Learning approaches to the simulation of mesoscale phenomena such as dislocations and morphological evolution of surfaces. The scope is methodological, and the Thesis tries to be self-contained. After an introduction to basic Machine Learning and Neural Networks concepts, different physical models are tackled. Methods are presented contextually to applications and are introduced from the simplest to the most sophisticated. Starting from the approximation of dislocation interactions in heteroepitaxial films, the Thesis proceeds in showing how to simulate the evolution of coherent films free surfaces and phase field models, using convolutional and recurrent NN respectively. Finally, an application of Generative Adversarial Networks to a stochastic mechanics problem is presented.

I metodi Machine Learning (ML) in generale e i Neural Network (NN) in particolare rappresentano una delle più attive aree di ricerca degli ultimi anni. La fisica e le Scienze dei Materiali computazionali non sono state immuni a questo sviluppo: approcci NN e ML stanno rapidamente venendo adottati come nuovi strumenti, con la promessa di poter risolvere problemi che fino ad ora sono rimasti intrattabili. In questa tesi verranno presentati alcuni sviluppi di questa direzione di ricerca, concentrandosi sulla possibilità di attaccare problemi tradizionali da nuovi punti di vista. In particolare, la discussione verterà come sia possibile adattare di metodi Machine Learning “puri” per la simulazione di fenomeni mesoscala, come dislocazioni e l’evoluzione morfologica di superfici. Lo scopo è metodologico e la Tesi cerca di essere auto sussistente. In seguito all’introduzione dei concetti generali del Machine Learning e dei Neural Network, si affrontano diversi modelli fisici. I metodi sono presentati contestualmente all’applicazione e sono presentati dal più semplice al più sofisticato. Iniziando con l’approssimazione dell’interazione tra dislocazioni in film eteroepitassiali con semplici feedforward fully connected Neural Netowork, la tesi si sviluppa quindi verso la simulazione dell’evoluzione delle superfici di film coerenti e modelli phase field, utilizzando metodi convoluzionali e ricorrenti rispettivamente. Infine, viene presentata una applicazione di Generative Adversarial Networks ad un problema di meccanica stocastica.

(2024). Deep Learning methods for the investigation of temporal evolution of materials. (Tesi di dottorato, Università degli Studi di Milano-Bicocca, 2024).

Deep Learning methods for the investigation of temporal evolution of materials

LANZONI, DANIELE
2024

Abstract

Machine Learning (ML) and Neural Networks (NN) specifically represent one of the most active areas of research in the last year. Computational physics and materials science have not been immune to this: Neural Networks and other ML approaches are being adopted as new tools, which promise to solve problems which have been intractable so far. In this Thesis, some advancements in this respect will be shown, focusing on the possibility of tackling traditional problems from new directions. In particular, the discussion will revolve around how to adapt pure Deep Learning approaches to the simulation of mesoscale phenomena such as dislocations and morphological evolution of surfaces. The scope is methodological, and the Thesis tries to be self-contained. After an introduction to basic Machine Learning and Neural Networks concepts, different physical models are tackled. Methods are presented contextually to applications and are introduced from the simplest to the most sophisticated. Starting from the approximation of dislocation interactions in heteroepitaxial films, the Thesis proceeds in showing how to simulate the evolution of coherent films free surfaces and phase field models, using convolutional and recurrent NN respectively. Finally, an application of Generative Adversarial Networks to a stochastic mechanics problem is presented.
MONTALENTI, FRANCESCO CIMBRO MATTIA
Modellizzazione; Neural Netowrks; Phase field; Cinetica; Mesoscale
Modeling; Neural Networks; Phase field; Kinetics; Mesoscale
FIS/03 - FISICA DELLA MATERIA
English
29-apr-2024
36
2022/2023
open
(2024). Deep Learning methods for the investigation of temporal evolution of materials. (Tesi di dottorato, Università degli Studi di Milano-Bicocca, 2024).
File in questo prodotto:
File Dimensione Formato  
phd_unimib_789367.pdf

accesso aperto

Descrizione: Deep Learning Methods for the Investigation of Temporal Evolution of materials
Tipologia di allegato: Doctoral thesis
Dimensione 13.89 MB
Formato Adobe PDF
13.89 MB Adobe PDF Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/10281/474580
Citazioni
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
Social impact