Date of Award

2024

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Mechanical Engineering

First Advisor

Jeffrey S. Marshall

Abstract

Mesoscale weather models and remote sensing methods are often tuned on datasets of microphysical properties of hydrometeors in cloud systems. Gaps in these datasets, especially those of snow, lead to qualitative and quantitative errors in weather prediction. Difficulty in data collection currently limits the size of available data. Image-based precipitation monitoring methods may help in data collection, allowing capture of falling snowflakes in high detail and allowing for estimation of their properties.

A novel Knowledge-Guided Convolutional Neural Network (KGCNN) is presented in this thesis for the prediction of various three-dimensional shape parameters, drag coefficient, mass, and density of falling snowflakes. The neural network model was pre-trained on silhouettes of synthetic snowflakes that are geometrically similar to real snowflakes, then fine-tuned on real snowflake images. Existing drag coefficient correlations were used to remove the burden of learning the relationship between the Reynolds number and drag coefficient from the model. The shape parameter outputs are constrained by custom physics-guided loss functions to ensure physical interpretability and allow for simultaneous prediction of particle volume and density. Pre-training and custom loss functions were found to reduce Normalized Root Mean Squared Error (NRMSE) on mass by 11.8%. Integration of drag coefficient correlations reduced mass NRMSE by 34.4% over similar models directly predicting drag coefficient, and outperformed existing physical correlations for snowflake drag coefficient. Predictions of shape parameters, volume, and density by the KGCNN were found to be consistent with experimental values.

As an example of an application of the KGCNN, videos of falling snow were taken using a GoPro Hero 11 Black with a 75 mm zoom lens shielded from wind. Shape parameters, drag coefficient, and mass are predicted for each sampled snowflake. The KGCNN can be used in this manner to predict properties of falling snow using snowflake images and measurements of falling velocity for scientific purposes.

Language

en

Number of Pages

91 p.

Available for download on Friday, April 04, 2025

Share

COinS