Skip to Main Content
 

Global Search Box

 
 
 
 

Files

File List


Mengfei_Yuan_Thesis_2019.pdf (5.49 MB)

ETD Abstract Container

Abstract Header

Machine Learning-Based Reduced-Order Modeling and Uncertainty Quantification for "Structure-Property" Relations for ICME Applications

Yuan, Mengfei
http://rave.ohiolink.edu/etdc/view?acc_num=osu1555580083945861

Abstract Details

2019, Doctor of Philosophy, Ohio State University, Materials Science and Engineering.
The design framework for complex materials property and processing models within the Integrated Computational Material Engineering (ICME) is often hindered by the expensive computational cost. The ultimate goal of ICME is to develop data-driven, materials-based tools for the concurrent optimization of material systems while, improving the deployment of innovative materials in real-world products. Reduced-order, fast-acting tools are essential for both bottom-up property prediction and top-down model calibrations employed for modern material design applications. Additionally, reduced-order modeling requires formal uncertainty quantification (UQ) from the processing stages all the way down to the manufacturing and component design. The goal of this thesis is to introduce a machine learning-based, reduced-order crystal plasticity model for face-centered cubic (FCC) polycrystalline materials. This implementation was founded upon Open Citrination, an open-sourced materials informatics platform. Case studies for both the bottom-up property prediction and top-down optimization of model parameters are demonstrated within this work. The proposed reduced-order model is used to correctly approximate the plastic stress-strain curves and the texture evolution under a range of deformation conditions and strain rates specific to a material. The inverted pathway is applied to quickly calibrate the optimal crystal plasticity hardening parameters given the macroscale stress-strain responses and evolutionary texture under certain processing conditions. A visco-plastic self-consistent (VPSC) method is used to create the training and validation datasets. The description of the material texture is given through a dimension reduction technique which was implemented by principal component analysis (PCA). The microstructures of engineering materials typically involve an intricate hierarchical crystallography, morphology, and composition. Therefore, an accurate, virtual representation of the microstructure is necessary for the prediction and evaluation of functional materials properties. Instead of analyzing larger amounts of experimental data, this work presents a state-of-the-art method which shows that the reliable virtual material texture can be generated using a convolutional neural network (CNN-based) generative modelling, generative adversarial networks (GANs). Moreover, uncertainty quantification is also critical to the reliability, safety, and quality of applications. The manufacturing uncertainties due to the microscale material variance usually produces unwanted effects on the macroscale components. Based off two different case studies, this thesis demonstrates the correlation between the material variance and the property uncertainty at hierarchical length scales, using a non-sampling method. One case study demonstrates that the uncertainty of energy absorption ability during crashing test due to the topological variance for pore composite material. The microstructural features are captured using statistical material descriptors (SMD), and the reduced-order “structure-property” maps are built upon the dimension reduction technique, kernel principal component analysis (KPCA). Polynomial chaos expansion (PCE) is applied to predict and evaluate the material randomness and the property uncertainty. Straightforward “microstructure - processing - property – performance” system is generated through sufficient sample-to-sample non-intrusive stochastic finite element (NISFE) simulations using probabilistic methods. Similarly, the other case study seeks to describe the “structure-property” relations for commercially pure titanium (CP-Ti) sheet metal during U-bending test. The goal is to show that controlling the variability of texture and the fraction coefficient can effectively minimize the springback effect during material processing.
Stephen Niezgoda (Advisor)
Yunzhi Wang (Committee Member)
Michael Groeber (Committee Member)
Maryam Ghazisaeidi (Committee Member)
139 p.
Materials Science
Machine learning, Reduced-order model, ICME, Microstructure, Generative adversarial networks, Uncertainty quantification, Polynomial Chaos, Stochastic computations, Finite element analysis

Recommended Citations

Citations

  • Yuan, M. (2019). Machine Learning-Based Reduced-Order Modeling and Uncertainty Quantification for "Structure-Property" Relations for ICME Applications [Doctoral dissertation, Ohio State University]. OhioLINK Electronic Theses and Dissertations Center. http://rave.ohiolink.edu/etdc/view?acc_num=osu1555580083945861

    APA Style (7th edition)

  • Yuan, Mengfei. Machine Learning-Based Reduced-Order Modeling and Uncertainty Quantification for "Structure-Property" Relations for ICME Applications. 2019. Ohio State University, Doctoral dissertation. OhioLINK Electronic Theses and Dissertations Center, http://rave.ohiolink.edu/etdc/view?acc_num=osu1555580083945861.

    MLA Style (8th edition)

  • Yuan, Mengfei. "Machine Learning-Based Reduced-Order Modeling and Uncertainty Quantification for "Structure-Property" Relations for ICME Applications." Doctoral dissertation, Ohio State University, 2019. http://rave.ohiolink.edu/etdc/view?acc_num=osu1555580083945861

    Chicago Manual of Style (17th edition)

osu1555580083945861
1,208
© 2019, all rights reserved.
This open access ETD is published by The Ohio State University and OhioLINK.