College of Science and Health Theses and Dissertations

Date of Award

Spring 6-11-2021

Degree Type

Thesis

Degree Name

Master of Science (MS)

Department

Physics

First Advisor

Eric Landahl, PhD

Second Advisor

Christopher Goedde, PhD

Third Advisor

Gabriela Gonzalez Aviles, PhD

Abstract

In soft-matter physics, Dynamic Light Scattering (DLS) stands out as the non-invasive technique designed to measure the size and distribution of particles using speckle correlation. Traditional DLS analysis performs a one-to-one mapping of the size with diffusion which is computationally costly, in addition to lacks in generalisation to seek deeper relationships amongst speckle patterns. A recent advancement in double-pulse speckle imaging allows for the temporal-correlation problem presented by DLS to be transformed into an image classification problem. In this prototype study, I probe the feasibility and efficacy of DLS data analysis of speckle correlation using Deep Learning (DL). To reduce the loss of generality and retain the integrity of the speckles, the study implements an image-focused DL algorithm. A Convolutional Neural Network (CNN or ConvNet) is used for feature extraction of pixelated images that are time-series in nature. Each scan consists of 500 frames (images) taken with double-pulses of 100 nanoseconds duration at 17 time delays. The CNN model was trained with roughly 1000 observations per sample with labelled target to supervise the training. The model picked up patterns within image matrices producing best performance accuracy of 83, average accuracy hovered at 58 percent. Treating speckle images to artificial intelligence to classify to the impurity type opens a new avenue to perform analysis in DLS. This prototype study shows that CNN could be an alternative candidate to traditional speckle image analysis.

SLP Collection

no

Download

Included in

Physics Commons

Share

COinS