Solid particles erosion is the phenomenon of material removal from structural components exposed to the impacts of solid particles transported by a fluid. This phenomenon is a significant concern for engineers, as it presents substantial challenges in maintaining the integrity and efficiency of systems and infrastructures. In order to predict the erosion of complex systems, a common approach consists in performing numerical simulations based on Computational Fluid Dynamics, in which an erosion model associates amount of material removal to every particle-wall impact. However, available erosion models rely on either a simplified treatment of the impact dynamics or on the fitting of predetermined regression models; as a result, they are often highly inaccurate. The idea of the research project of which this thesis is the first step is to develop an erosion model whose formulation is obtained directly from the data through Machine Learning methods. In meeting this challenge, this thesis delves into the fundamentals of Machine Learning, with a particular focus on supervised learning methods applied to erosion prediction. Through the analysis of experimental data from Dry Direct Impact tests reported in the literature, and by focusing on the impact of the two key fluid dynamic variables, namely particle impact velocity and particle impact angles, datasets are constructed to develop and validate predictive models. The study uses analysis of different types of Machine Learning methods (simple regression, polynomial regression, and multivariate regression) to identify relationships between erosion ratios and influential parameters. During the process, the importance of data quality, scalability and representativeness is explored, addressing the problems associated with the limited availability and uncertainty of experimental erosion data. By applying regression techniques such as gradient descent method, regularization and k-fold cross-validation, the research aims to assess the reliability and provide guidelines for the development of Machine Learning based erosion models, presenting a good base for future work.
L'erosione di particelle solide rappresenta il fenomeno di rimozione di materiale da componenti strutturali esposti all'impatto di particelle solide trasportate da un fluido. Questo fenomeno implica sfide sostanziali per gli ingegneri nel mantenere l'integrità e l'efficienza di infrastrutture. Per prevedere l'erosione di sistemi complessi, un approccio comune consiste nell'eseguire simulazioni numeriche basate sulla Fluidodinamica Computazionale, in cui un modello di erosione associa la quantità di materiale rimosso a ogni impatto tra particelle e pareti. Tuttavia, i modelli di erosione disponibili si basano su un trattamento semplificato delle dinamiche di impatto o sull'adattamento di modelli di regressione predeterminati; di conseguenza, sono spesso inaccurati. L'idea della tesi è quella di sviluppare un modello di erosione la cui formulazione sia ottenuta direttamente dai dati attraverso metodi di Apprendimento Automatico. Per affrontare questa sfida, questa tesi approfondisce i fondamenti dell’Apprendimento Automatico, con particolare attenzione ai metodi di apprendimento supervisionato. Attraverso l'analisi di dati sperimentali provenienti da Test di Impatto Diretto a Secco riportati in letteratura, e concentrandosi sull'impatto delle due variabili fluidodinamiche chiave, velocità di impatto delle particelle e angoli di impatto delle particelle, vengono costruiti set di dati per sviluppare e validare modelli predittivi. Lo studio utilizza l'analisi di diversi tipi di metodi di Apprendimento Automatico (regressione semplice, regressione polinomiale e regressione multivariata) per identificare le relazioni tra tassi di erosione e parametri influenti. Durante il processo, si esplora l'importanza della qualità, della scalabilità e della rappresentatività dei dati, affrontando i problemi associati alla limitata disponibilità e all'incertezza dei dati sperimentali sull'erosione. Applicando tecniche di regressione come il metodo di discesa del gradiente, la regolarizzazione e la convalida incrociata k-fold, la ricerca mira a valutare l'affidabilità e a fornire linee guida per lo sviluppo di modelli di erosione basati sull’Apprendimento Automatico, presentando una buona base per il lavoro futuro.
Exploration of machine learning methods for the prediction of solid particle erosion
Martinelli, Yuri
2023/2024
Abstract
Solid particles erosion is the phenomenon of material removal from structural components exposed to the impacts of solid particles transported by a fluid. This phenomenon is a significant concern for engineers, as it presents substantial challenges in maintaining the integrity and efficiency of systems and infrastructures. In order to predict the erosion of complex systems, a common approach consists in performing numerical simulations based on Computational Fluid Dynamics, in which an erosion model associates amount of material removal to every particle-wall impact. However, available erosion models rely on either a simplified treatment of the impact dynamics or on the fitting of predetermined regression models; as a result, they are often highly inaccurate. The idea of the research project of which this thesis is the first step is to develop an erosion model whose formulation is obtained directly from the data through Machine Learning methods. In meeting this challenge, this thesis delves into the fundamentals of Machine Learning, with a particular focus on supervised learning methods applied to erosion prediction. Through the analysis of experimental data from Dry Direct Impact tests reported in the literature, and by focusing on the impact of the two key fluid dynamic variables, namely particle impact velocity and particle impact angles, datasets are constructed to develop and validate predictive models. The study uses analysis of different types of Machine Learning methods (simple regression, polynomial regression, and multivariate regression) to identify relationships between erosion ratios and influential parameters. During the process, the importance of data quality, scalability and representativeness is explored, addressing the problems associated with the limited availability and uncertainty of experimental erosion data. By applying regression techniques such as gradient descent method, regularization and k-fold cross-validation, the research aims to assess the reliability and provide guidelines for the development of Machine Learning based erosion models, presenting a good base for future work.File | Dimensione | Formato | |
---|---|---|---|
2024_04_Martinelli.pdf
solo utenti autorizzati dal 20/03/2025
Dimensione
7.83 MB
Formato
Adobe PDF
|
7.83 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/10589/219481