서브메뉴
검색
Interpretable Design of Microstructural Material Systems With Mixed-Variable Representations
Interpretable Design of Microstructural Material Systems With Mixed-Variable Representations
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20250211151409
- ISBN
- 9798382761800
- DDC
- 621
- 서명/저자
- Interpretable Design of Microstructural Material Systems With Mixed-Variable Representations
- 발행사항
- [Sl] : Northwestern University, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 169 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 85-11, Section: B.
- 주기사항
- Advisor: Chen, Wei.
- 학위논문주기
- Thesis (Ph.D.)--Northwestern University, 2024.
- 초록/해제
- 요약The intersection of engineering design and material science has led to the proliferation of materials design research. In the relentless pursuit of technological excellence, the design and development of engineered material systems lay the foundation for groundbreaking discoveries and transformative applications. With the advancements in computational modeling, optimization, and evaluation capabilities, the development of engineered material systems has taken a rapid acceleration. On the other hand, due to the complexity of these computational innovations, the understanding and interpretability behind the development of material systems have become a challenge that needs to be addressed. Within this consideration, interpretable materials design research area has emerged to decompose the complex relationships between the material ecosystems from the perspective of engineering design. Specifically, interpretable materials design research can be defined as the process of purposefully making data-driven design decisions that captures the cause-effect relationships between the material ecosystem for building physics informed process-structure-property links to discover scientific knowledge about the material system. In the quest for contributing to interpretable material design research, the main theme of this dissertation is the development of design methodologies and frameworks through machine learning and statistical techniques to extract scientific knowledge and interpretability for material systems with different material representation requirements. Through this theme, this dissertation is dedicated to addressing four identified challenges that arise from different perspectives and domains of interpretable materials design research.The first research challenge emerges when material systems require or possess qualitative information. Although most qualitative design variables are easily accessible, it is very rare that only qualitative variables are utilized to design material systems. Furthermore, the design space can easily expand towards high order magnitudes with the combination of different variables, requiring high amount of resource allocation to identify novel designs. In this dissertation, a Latent Variable Gaussian Process (LVGP) and multi-objective batch Bayesian optimization framework is developed to identify top-performing material candidates from a large combinatorial design space adaptively, autonomously, and efficiently. The contribution of this work is a design framework that requires no specific physical descriptors for global optimization of large combinatorial material spaces while providing interpretability through latent variables with physical justification.The second challenge arises from the lack of extracting the importance and interactions within the mixed-variable material design spaces on material properties. For quantitative spaces, a well-known method to extract such information is achieved through global sensitivity analysis (GSA) methods. Thus far, GSA techniques are limited to studies with only quantitative design variables, even though qualitative design variables are ubiquitous in many materials design applications. In this dissertation, a new sampling approach is integrated into metamodeling to develop a mixed-variable GSA methodology. The contribution of this work lies in incorporating qualitative variables into GSA studies and further integrating it with Bayesian optimization to create a sensitivity-aware design framework.The success of interpretable materials design relies heavily on the availability and quality of materials information (data). As a result, the third challenge stems from the sparse and possibly unreliable data resulting from inconsistencies and discrepancies in the data curation methodologies across various information sources, leading to unavoidable difficulties that can have severe impact on subsequent data-driven analysis, modeling, and interpretation of the material system. In this dissertation, the inherent complexities and uncertainties associated with each information source are extracted and further implemented into data fusion modeling through incorporation of information each source as a design variable into multi-source modeling through LVGP to obtain comprehensive, robust, and deeper understanding of material behaviors. The contribution of this work comes from leveraging the complementary information available from and in between multiple sources to enable the development of more interpretable and impactful modeling of material systems.Finally, under different processing and experimental conditions, materials go through multiscale changes in their structures, affecting their properties and performance significantly. Consequently, data-driven materials design is subject to numerous sources of uncertainty at each length scale of the material system that impedes the understanding and the pace of the materials development. Therefore, uncertainty quantification (UQ) and propagation (UP) play a vital role for generating safe design guidelines for materials systems. However, performing UQ & UP in multiscale materials systems often requires computationally intensive simulations and analyses, particularly when considering high-dimensional parameter spaces and complexity of material systems. In this dissertation, an efficient multiscale UQ & UP framework through copula-based sampling is developed to significantly reduce the amount of resources required to understand and extract the influence of lower scale uncertainties on the highest material scale properties.
- 일반주제명
- Mechanical engineering
- 일반주제명
- Materials science
- 키워드
- Materials design
- 기타저자
- Northwestern University Mechanical Engineering
- 기본자료저록
- Dissertations Abstracts International. 85-11B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008250123s2024 us c eng d■001000017161535
■00520250211151409
■006m o d
■007cr#unu||||||||
■020 ▼a9798382761800
■035 ▼a(MiAaPQ)AAI31292473
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a621
■1001 ▼aComlek, Yigitcan.▼0(orcid)0000-0002-3654-1576
■24510▼aInterpretable Design of Microstructural Material Systems With Mixed-Variable Representations
■260 ▼a[Sl]▼bNorthwestern University▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a169 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 85-11, Section: B.
■500 ▼aAdvisor: Chen, Wei.
■5021 ▼aThesis (Ph.D.)--Northwestern University, 2024.
■520 ▼aThe intersection of engineering design and material science has led to the proliferation of materials design research. In the relentless pursuit of technological excellence, the design and development of engineered material systems lay the foundation for groundbreaking discoveries and transformative applications. With the advancements in computational modeling, optimization, and evaluation capabilities, the development of engineered material systems has taken a rapid acceleration. On the other hand, due to the complexity of these computational innovations, the understanding and interpretability behind the development of material systems have become a challenge that needs to be addressed. Within this consideration, interpretable materials design research area has emerged to decompose the complex relationships between the material ecosystems from the perspective of engineering design. Specifically, interpretable materials design research can be defined as the process of purposefully making data-driven design decisions that captures the cause-effect relationships between the material ecosystem for building physics informed process-structure-property links to discover scientific knowledge about the material system. In the quest for contributing to interpretable material design research, the main theme of this dissertation is the development of design methodologies and frameworks through machine learning and statistical techniques to extract scientific knowledge and interpretability for material systems with different material representation requirements. Through this theme, this dissertation is dedicated to addressing four identified challenges that arise from different perspectives and domains of interpretable materials design research.The first research challenge emerges when material systems require or possess qualitative information. Although most qualitative design variables are easily accessible, it is very rare that only qualitative variables are utilized to design material systems. Furthermore, the design space can easily expand towards high order magnitudes with the combination of different variables, requiring high amount of resource allocation to identify novel designs. In this dissertation, a Latent Variable Gaussian Process (LVGP) and multi-objective batch Bayesian optimization framework is developed to identify top-performing material candidates from a large combinatorial design space adaptively, autonomously, and efficiently. The contribution of this work is a design framework that requires no specific physical descriptors for global optimization of large combinatorial material spaces while providing interpretability through latent variables with physical justification.The second challenge arises from the lack of extracting the importance and interactions within the mixed-variable material design spaces on material properties. For quantitative spaces, a well-known method to extract such information is achieved through global sensitivity analysis (GSA) methods. Thus far, GSA techniques are limited to studies with only quantitative design variables, even though qualitative design variables are ubiquitous in many materials design applications. In this dissertation, a new sampling approach is integrated into metamodeling to develop a mixed-variable GSA methodology. The contribution of this work lies in incorporating qualitative variables into GSA studies and further integrating it with Bayesian optimization to create a sensitivity-aware design framework.The success of interpretable materials design relies heavily on the availability and quality of materials information (data). As a result, the third challenge stems from the sparse and possibly unreliable data resulting from inconsistencies and discrepancies in the data curation methodologies across various information sources, leading to unavoidable difficulties that can have severe impact on subsequent data-driven analysis, modeling, and interpretation of the material system. In this dissertation, the inherent complexities and uncertainties associated with each information source are extracted and further implemented into data fusion modeling through incorporation of information each source as a design variable into multi-source modeling through LVGP to obtain comprehensive, robust, and deeper understanding of material behaviors. The contribution of this work comes from leveraging the complementary information available from and in between multiple sources to enable the development of more interpretable and impactful modeling of material systems.Finally, under different processing and experimental conditions, materials go through multiscale changes in their structures, affecting their properties and performance significantly. Consequently, data-driven materials design is subject to numerous sources of uncertainty at each length scale of the material system that impedes the understanding and the pace of the materials development. Therefore, uncertainty quantification (UQ) and propagation (UP) play a vital role for generating safe design guidelines for materials systems. However, performing UQ & UP in multiscale materials systems often requires computationally intensive simulations and analyses, particularly when considering high-dimensional parameter spaces and complexity of material systems. In this dissertation, an efficient multiscale UQ & UP framework through copula-based sampling is developed to significantly reduce the amount of resources required to understand and extract the influence of lower scale uncertainties on the highest material scale properties.
■590 ▼aSchool code: 0163.
■650 4▼aMechanical engineering
■650 4▼aMaterials science
■653 ▼aEngineering design
■653 ▼aInterpretable machine learning
■653 ▼aMaterials design
■653 ▼aUncertainty quantification
■653 ▼aLatent Variable Gaussian Process
■690 ▼a0548
■690 ▼a0800
■690 ▼a0794
■71020▼aNorthwestern University▼bMechanical Engineering.
■7730 ▼tDissertations Abstracts International▼g85-11B.
■790 ▼a0163
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17161535▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


