Graduation Semester and Year
Spring 2026
Document Type
Dissertation
Degree Name
Doctor of Philosophy in Mechanical Engineering
Department
Mechanical and Aerospace Engineering
First Advisor
Paul Davidson
Second Advisor
Endel Iarve
Third Advisor
Xin Liu
Fourth Advisor
Shiyao Lin
Fifth Advisor
Emma Yang
Abstract
Composite laminates are highly sought after in the aerospace industry as they provide strength without dramatically increasing the weight of manufactured structures. However, composite laminates are susceptible to barely visible impact damage caused by routine activities. This type of damage can easily go unnoticed while significantly reducing the load-carrying capability of the laminate. Current non-destructive evaluation techniques, such as ultrasonic scanning, can reveal the damage footprint but provide no insight into delamination through-the-thickness due to the shadowing effect. Micro-computed tomography offers ply-by-ply damage resolution but is unsuitable for field inspections and is constrained by specimen size. This study aims to reconstruct damage morphologies, typically only visible through destructive inspection or micro-computed tomography, from sparse ultrasonic scans, using a double UNet machine learning model trained on numerical simulation data.
The first portion of this dissertation focuses on the development of an automated tool for generating impact models capable of reproducing experimentally observed damage morphologies using discrete damage modeling techniques. This tool enables the efficient creation of laminate models across multiple configurations, significantly reducing model development time. The generated models are used to investigate the effects of compliant boundary conditions on damage morphology in laminates with varying length-to-thickness ratios. The results indicate that compliant boundaries influence the force–displacement response when damage is absent, but have negligible effect once damage is incorporated. Additionally, the appropriate crack spacing required to capture different damage morphologies is identified; a spacing of 1.5 mm is found to limit delamination growth and promote the formation of delamination "petals." Finally, the influence of cohesive parameters in crack cohesive elements is examined, establishing their relationship with both the force–displacement response and the resulting delamination morphology.
The second focus is on data generation using the automated tool to produce damage morphology results for laminates with varying stack-up configurations. A method to extract data from numerical simulations and integrate it into a machine learning framework is developed, designed to be compatible with experimental ultrasonic scan formats. A double UNet pipeline is then constructed and evaluated using synthetic data resembling experimental scans. The model successfully segments visible damage from pseudo-CT scans; however, it struggles with the reconstruction step in the second UNet, likely due to insufficient spatial context in the training data. Together, these contributions represent a step toward a practical framework for reconstructing full damage morphologies from non-destructive ultrasonic measurements, with implications for structural health monitoring in aerospace applications.
Keywords
Composites, Low velocity impact, Barely visible impact damage, Machine learning, non destructive inspection, U-Net, progressive damage modeling, delamination morphology, cohesive elements
Disciplines
Applied Mechanics | Other Mechanical Engineering
License
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License
Recommended Citation
Valdez, Oscar A., "NUMERICAL AND MACHINE LEARNING BASED RECREATION OF DAMAGE MORPHOLOGIES OF BARELY VISIBLE IMPACT DAMAGE" (2026). Mechanical and Aerospace Engineering Dissertations. 2.
https://mavmatrix.uta.edu/mechaerospace_dissertations2/2