Thesis
Machine-learning-based multiscale modeling of spatially tailored materials via peridynamics
University of Iowa
Master of Science (MS), University of Iowa
Spring 2021
DOI: 10.17077/etd.005853
Abstract
Advancements in technology have always been limited by the availability of the necessary materials. The greatest marvels in engineering have both caused and motived advances in materials science. This symbiotic relationship between advancements in science and technology as well as advancements in materials science can be observed throughout history. Man would not have been able to take its first steps on the moon had the materials to make rockets not existed, while the materials for the rockets would not have existed if man had not wished to step on the moon. Studying and detailing the strengths and weaknesses of different categories of materials as well as identifying their behavior is key to the future advancement of science and engineering.An interesting class of materials that is gaining attention in engineering today are functionally graded materials (FGMs), which are a class of spatially tailored materials that aim to create composite materials that have volume fractions which vary with respect to space. Essentially, they are a class of materials which takes the benefits of multiple materials and combine them into one optimized version which determines which material benefits are needed in certain regions of the material. Imagine that a special material is needed for an application in which one side of the material will be subject to extremely hot temperatures, yet it must be resistant to deformation. Hot temperatures weaken metals, but those same temperatures do not affect ceramics. Similarly, metals are much more resistant to catastrophic material failure than ceramics. An FGM would aim to design a material using both metal and ceramic with a specified volume fraction distribution along the cross-section to maximize the effects of both.
The main issue with FGM systems is that the material properties have inherent randomness in their material property values in real world applications due in part to the methods used to create them. This implies that a sample FGM could have certain section with slightly higher volume in a than a second sample. These random effects are also expressed at the microscale because of microstructural irregularities (i.e. changes in local volume fraction or small pores). There is great need for a model that accurately describes and compiles these random effects.
Machine learning models offer a potential avenue for research in this manner. With the recent rise of machine learning, it is possible to train a machine to recognize patterns in the behavior of the material properties of FGMs. To do this, a large set of data will be required for the machine to obtain an accurate reproduction of the innate patterns found within the material properties of FGMs. Generating that large set of data also poses its own challenges.
Peridynamics is a continuum material model which has great benefits to one who wishes to make a material model that accurately reproduces the effects of classical solid mechanics. Having a model which reproduces these effects at the microscale would become a powerful foundation for a dataset which can be used to train a machine. The work in this thesis explores the effects of three material properties at the microscale, the process of training a Bayesian machine learning algorithm to recognize patterns within the dataset, and ultimately the use of those patterns to design inherently randomized FGM material models.
The randomized FGM models are macroscale models that will be created using a machine learning algorithm which was trained using the data obtained from the microscale model. The microstructural abnormalities commonly found in FGMs will be communicated to the macroscale through the machine learning algorithm when creating the macroscale models. The configurations are then used in two fracture analysis simulations (mode I fracture and plate-with-a-hole fracture propagation).
Details
- Title: Subtitle
- Machine-learning-based multiscale modeling of spatially tailored materials via peridynamics
- Creators
- Ahmed ElTuhami
- Contributors
- Shaoping Xiao (Advisor)Caterina Lamuta (Committee Member)Phillip Deierling (Committee Member)
- Resource Type
- Thesis
- Degree Awarded
- Master of Science (MS), University of Iowa
- Degree in
- Mechanical Engineering
- Date degree season
- Spring 2021
- DOI
- 10.17077/etd.005853
- Publisher
- University of Iowa
- Number of pages
- xiv, 117 pages
- Copyright
- Copyright 2021 Ahmed ElTuhami
- Language
- English
- Description illustrations
- illustrations (some color)
- Description bibliographic
- Includes bibliographical references (pages 116-117)
- Public Abstract (ETD)
The greatest marvels in engineering have both caused and motived advances in materials science. This symbiotic relationship between advancements in science and technology as well as advancements in materials science can be observed throughout history. Man would not have been able to take its first steps on the moon had the materials to make rockets not existed, while the materials for the rockets would not have existed if man had not wished to step on the moon. Studying and detailing the strengths and weaknesses of different categories of materials as well as identifying their behavior is key to the future advancement of science and engineering.
An interesting class of materials that is gaining attention in engineering today are functionally graded materials (FGMs). Essentially, they are a class of materials which takes the benefits of multiple materials and combine them into one optimized version which determines which material benefits are needed in certain regions of the material. Imagine that a special material is needed for an application in which one side of the material will be subject to extremely hot temperatures, yet it must be resistant to deformation. Such applications are the driving force for the recent rise in FGMs.
The main issue with FGM systems is that the material properties have inherent randomness in their material property values in real world applications due in part to the methods used to create them. There is great need for a model that accurately describes and compiles these random effects. With the recent rise of machine learning, it is possible to train a machine to recognize patterns in the behavior of the material properties of FGMs. The work in this thesis explores the effects of three material properties at the microscale, the process of training a Bayesian machine learning algorithm to recognize patterns within the dataset, and ultimately the use of those patterns to design inherently randomized FGM material models.
- Academic Unit
- Mechanical Engineering
- Record Identifier
- 9984097276302771
Metrics
37 File views/ downloads
114 Record Views