Deep learning enabled materials design and characterization
No Thumbnail Available
Authors
Meeting name
Sponsors
Date
Journal Title
Format
Thesis
Subject
Abstract
In this dissertation, deep learning methodologies are applied to innovative computational approaches for the development and characterization of materials in the material science field. The first part of the research focuses on a physics-informed machine learning workflow for virtual experimentation in the 3D printing of thermoplastics, where traditional methods are limited by complex chemical reactions and extensive design possibilities. Utilizing a dataset of 62 formulations and 216 Stress-Strain curves, this method employs dimension reduction and a novel machine learning model with physics-informed descriptors to simulate over 100,000 virtual experiment sets in under one minute, significantly enhancing the speed of material discovery. The second part improves the analysis of characterization data, specifically X-ray diffraction (XRD) patterns, by using Transformer-based models that surpass previous CNN-based models in training speed and accuracy. This segment introduces a novel data augmentation technique that simulates experimental errors and uses interpretability analysis to show how the model captures long-distance interactions between XRD peaks. It also explores the potential of transfer learning from XRD to Fourier-transform infrared spectroscopy (FTIR) data, broadening the model's applicability and improving the efficiency and accuracy of material characterization. Overall, this dissertation demonstrates how deep learning can revolutionize material science research, providing faster, more accurate tools for material development and analysis.
Table of Contents
DOI
PubMed ID
Degree
Ph. D.