Date of Award

2024

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Complex Systems and Data Science

First Advisor

Safwan Wshah

Abstract

The drag coefficient of snowflakes is an crucial particle descriptor that can quantify the relationships with the mass, shape, size, and fall speed of snowflake particles. Previous studies has relied on estimating and improving empirical correlations for the drag coefficient of particles, utilizing 3D images from the Multi-Angled Snowflake Camera Database (MASCDB) to estimate snowflake properties such as mass, geometry, shape classification, and rimming degree. However, predictions of the drag coefficient with single-view 2D images of snowflakes has proven to be a challenging problem, primarily due to the lack of data and time-consuming, expensive methods used to estimate snowflake shape factors such as sphericity and convex hull. In this paper, we propose a cost-effective and time-efficient approach to address the challenges in predicting the drag coefficients from single-view 2D images of falling snowflakes. Our method combines EfficientNetB7 for image preprocessing to remove the background and border from snowflake images, Kernel Principal Component Analysis (KPCA) to extract meaningful features from the snowflake images, and Machine Learning methods, namely Random Forests, XGBoost models, Multilayer Perceptron (MLP) models, and MLP models trained on distinct Reynolds number flow regimes, to predict drag coefficients using the Locatelli and Hobbs dataset. Through comprehensive evaluation, our model achieved a mean squared error of 0.195, outperforming most existing empirical correlations. Moreover, an evaluation of the feature importance using mean decrease impurity (MDI) showed that the KPCA feature extraction added influential and meaningful data points to our machine learning models.

Language

en

Number of Pages

70 p.

Download

Share

COinS