In the present thesis a Machine Learning approach in Fluid Mechanics field was investigated. In particular Artificial Neural Networks were used to predict lift and drag coefficient of NACA 4 digit airfoils. In the last years the application of Artificial Intelligence and in particular of Machine Learning to scientific disciplines increased substantially. Machine Learning offers techniques to extract information and knowledge from data and it provides the possibility to handle with massive quantitative of data. The purpose of the related work was to investigate how Machine Learning is working, in particular Neural Networks, and how it has to be applied in order to make an aerodynamic prediction. Due to the aim of this work the predictions of lift and drag coefficient was considered as study case. This study case gave the opportunity to investigate NN for making an aerodynamic prediction and it leads to have a new tool for making drag and lift predictions. The preliminary phase of the work was to create the data-set necessary for the secondary phase, the Neural Network analysis. The generation of the dataset involved CFD simulations. Those were performed with DLR-TAU code, a finite volume method for RANS equations. A tool as a code for RANS equations were used because of its ability to capture the aerodynamic coefficients of interest in the related work. The preliminary phase includes also all the steps that a CFD simulation concerns: CAD generation performed with Geocreate, mesh generation with Centaur and numerical simulation with DLR-TAU code. The related theoretical background is given in chapter 1, in chapter 3 instead numerical simulations are presented. In the context of neural network approach the software package Google TensorFlow 2 via Python3 interface was used. Therefore, in the context of this thesis, artificial neural networks were used to manage lift and drag coefficient generated from CFD simulations. This work presents in chapter 2 an overview of what is Machine Learning and a detailed introduction to artificial neural networks. Final results and considerations are shown in chapter 4.
Nella presente tesi è stato investigato l’utilizzo dell’apprendimento automatico (noto anche come Machine Learning) in campo fluidodinamico. In particolare, le reti neurali artificiali sono state considerate per predire il coefficiente di portanza e il coefficiente di resistenza per profili appartenenti alla serie NACA 4-digit. Negli ultimi anni l’utilizzo dell’apprendimento automatico e in particolare delle reti neurali artificiali è cresciuto notevolmente. L’apprendimento automatico prevede approcci che forniscono la possibilità di estrarre informazioni dai dati di interesse e permette anche la gestione di elevate quantità di dati. L’obiettivo di questo elaborato era quello di investigare i principi fondamentali dell’apprendimento automatico, in particolare delle reti neurali artificiali, e come debbano essere applicate per generare una previsione aerodinamica. Considerato l’obiettivo dell’elaborato, la predizione di resistenza e portanza è il caso studio coinvolto. Questo caso studio fornisce l’opportunità di investigare le reti neurali artificiali quando vengono utilizzate per una predizione aerodinamica e fornisce anche l’opportunità di sviluppare un nuovo strumento per la predizione di portanza e resistenza. La fase preliminare ha compreso la generazione dei dati necessari per la seconda fase, l’analisi tramite le reti neurali artificiali. Le generazione dei dati necessari ha coinvolto simulazioni CFD. Le ultime state eseguite tramite il codice DLR-TAU, un metodo numerico ai volumi finiti per le equazioni RANS. Uno strumento come un codice per le RANS è stato preso in considerazione per le la sua abilità nel ”catturare” i coefficienti aerodinamici di interesse in questo progetto. La fase preliminare include anche tutti i preparativi per una simulazione CFD: la generazione del modello, il software Geocreate è stato utilizzato a tale scopo; la generazione della griglia, effettuata con Centaur ed infine la simulazione numerica, svolta con il codice DLR-TAU. Le relative conoscenze teoriche sono fornite nel capitolo 1, invece nel capitolo 3 le simulazioni numeriche sono presentate. Riguardo all’approccio con le reti neurali è stato utilizzata la libreria software TensorFlow 2 di Google tramite interfaccia Python3. Dunque, in questo progetto di tesi le reti neurali sono utilizzate per gestire i coefficienti di portanza e di resistenza generati tramite le simulazioni CFD. Nel capitolo 2 sono introdotti un riepilogo dell’apprendimento automatico e una dettagliata descrizione delle reti neurali artificiali. I risultati e le considerazioni finali sono presentati nel capitolo 4.
A neural network approach for predicting the arerodynamic performance of airfoils
SPINI, ANDREA
2019/2020
Abstract
In the present thesis a Machine Learning approach in Fluid Mechanics field was investigated. In particular Artificial Neural Networks were used to predict lift and drag coefficient of NACA 4 digit airfoils. In the last years the application of Artificial Intelligence and in particular of Machine Learning to scientific disciplines increased substantially. Machine Learning offers techniques to extract information and knowledge from data and it provides the possibility to handle with massive quantitative of data. The purpose of the related work was to investigate how Machine Learning is working, in particular Neural Networks, and how it has to be applied in order to make an aerodynamic prediction. Due to the aim of this work the predictions of lift and drag coefficient was considered as study case. This study case gave the opportunity to investigate NN for making an aerodynamic prediction and it leads to have a new tool for making drag and lift predictions. The preliminary phase of the work was to create the data-set necessary for the secondary phase, the Neural Network analysis. The generation of the dataset involved CFD simulations. Those were performed with DLR-TAU code, a finite volume method for RANS equations. A tool as a code for RANS equations were used because of its ability to capture the aerodynamic coefficients of interest in the related work. The preliminary phase includes also all the steps that a CFD simulation concerns: CAD generation performed with Geocreate, mesh generation with Centaur and numerical simulation with DLR-TAU code. The related theoretical background is given in chapter 1, in chapter 3 instead numerical simulations are presented. In the context of neural network approach the software package Google TensorFlow 2 via Python3 interface was used. Therefore, in the context of this thesis, artificial neural networks were used to manage lift and drag coefficient generated from CFD simulations. This work presents in chapter 2 an overview of what is Machine Learning and a detailed introduction to artificial neural networks. Final results and considerations are shown in chapter 4.File | Dimensione | Formato | |
---|---|---|---|
2020_06_SPINI.pdf
accessibile in internet per tutti
Descrizione: Tesi Andrea Spini
Dimensione
6.63 MB
Formato
Adobe PDF
|
6.63 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/10589/164519