Analysis of Composite Material Properties with the Reactive Interface Force Field and Hierarchical Spatial Graph Neural Networks
Analysis of Composite Material Properties with the Reactive Interface Force Field and Hierarchical Spatial Graph Neural Networks
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20250211151144
- ISBN
- 9798382715537
- DDC
- 542
- 서명/저자
- Analysis of Composite Material Properties with the Reactive Interface Force Field and Hierarchical Spatial Graph Neural Networks
- 발행사항
- [Sl] : University of Colorado at Boulder, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 173 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 85-11, Section: B.
- 주기사항
- Advisor: Heinz, Hendrik.
- 학위논문주기
- Thesis (Ph.D.)--University of Colorado at Boulder, 2024.
- 초록/해제
- 요약The failure mechanisms of materials at different length scales, from the atomic to the macroscopic scale, pose a grand challenge for applications in biomaterials, aerospace, construction, and commodity applications. Chemistry, defects, entanglements of polymers in composites, the organization of nanostructures, and the multi-scale dynamics affect properties such as modulus, strength, ductility, and toughness. Designing materials with desired mechanical properties and other functionality is difficult and time consuming, even with modern experimental and computational methods. Molecular dynamics (MD) simulations can help solve this problem by providing atomic-level information on mechanisms and mechanical properties at the nanoscale, which can be hard to measure experimentally. MD simulations can also speed up the process of identifying and developing new materials by exploring different compositions, assemblies, chemical reactions, and processing conditions, compared to other computational methods and experimental setups. This thesis advances the development of methods to simulate chemical reactions, introducing the reactive INTERFACE force field (IFF-R), utilizes IFF-R to create a database of over 1000 models of defective nanoscale graphitic structures and associated stress-strain curves, and introduces physics-informed machines learning methods to accelerate mechanical property predictions relative to MD simulations. First, improvements in modeling graphite and carbon nanotubes (CNT) by including virtual π-electrons using the Interface force field (IFF) are reported, including in-depth validation relative to experimental data, including 12-6 and 9-6 options of Lennard-Jones potentials. The Reactive Interface Force Field (IFF-R) is introduced, which allows the simulation of bond breaking using a substitution of harmonic bond stretching potentials with Morse bond potentials, maintains the high accuracy and speed of IFF, and eliminates discontinuities in bond energy upon dissociation. IFF-R also allows continuous simulations where bonds can break and form without needing a complex bond-order function, for example, using a combination with template-based reaction frameworks such as REACTER in LAMMPS. IFF-R is applicable to metals, ceramics, composites, and biopolymers such as cellulose and proteins. MD simulations with IFF-R were further used to create a dataset of over 1000 structures of pristine and defective CNT bundles and graphitic morphologies up to 80,000 atoms, as well as the associated stress-strain curves up to failure. The data set was used to develop and test machine-learning (ML) algorithms to predict the tensile modulus and strength for CNT assemblies, CNT junctions, and carbon-fiber cross-sections from the 3D atomic structure in up to about 1,000 faster speed than MD simulations. The most reliable ML method, called the hierarchical spatial graph neural network (HS-GNN), contains spatial information on bonds and torsions, as well as multi-scale resolution to include long-range interactions. HS-GNN performs well for sparse data and topologies outside the training range. The ML methods can be used to quickly screen computer generated 3D graphitic models for best mechanical properties to inform choices for composite materials, and can be extended to other chemistries and nanostructures.
- 일반주제명
- Computational chemistry
- 일반주제명
- Materials science
- 키워드
- Carbon
- 키워드
- Composites
- 키워드
- Force field
- 키워드
- Machine learning
- 기타저자
- University of Colorado at Boulder Materials Science and Engineering
- 기본자료저록
- Dissertations Abstracts International. 85-11B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008250123s2024 us c eng d■001000017160974
■00520250211151144
■006m o d
■007cr#unu||||||||
■020 ▼a9798382715537
■035 ▼a(MiAaPQ)AAI31234741
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a542
■1001 ▼aWinetrout, Jordan Jacob.
■24510▼aAnalysis of Composite Material Properties with the Reactive Interface Force Field and Hierarchical Spatial Graph Neural Networks
■260 ▼a[Sl]▼bUniversity of Colorado at Boulder▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a173 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 85-11, Section: B.
■500 ▼aAdvisor: Heinz, Hendrik.
■5021 ▼aThesis (Ph.D.)--University of Colorado at Boulder, 2024.
■520 ▼aThe failure mechanisms of materials at different length scales, from the atomic to the macroscopic scale, pose a grand challenge for applications in biomaterials, aerospace, construction, and commodity applications. Chemistry, defects, entanglements of polymers in composites, the organization of nanostructures, and the multi-scale dynamics affect properties such as modulus, strength, ductility, and toughness. Designing materials with desired mechanical properties and other functionality is difficult and time consuming, even with modern experimental and computational methods. Molecular dynamics (MD) simulations can help solve this problem by providing atomic-level information on mechanisms and mechanical properties at the nanoscale, which can be hard to measure experimentally. MD simulations can also speed up the process of identifying and developing new materials by exploring different compositions, assemblies, chemical reactions, and processing conditions, compared to other computational methods and experimental setups. This thesis advances the development of methods to simulate chemical reactions, introducing the reactive INTERFACE force field (IFF-R), utilizes IFF-R to create a database of over 1000 models of defective nanoscale graphitic structures and associated stress-strain curves, and introduces physics-informed machines learning methods to accelerate mechanical property predictions relative to MD simulations. First, improvements in modeling graphite and carbon nanotubes (CNT) by including virtual π-electrons using the Interface force field (IFF) are reported, including in-depth validation relative to experimental data, including 12-6 and 9-6 options of Lennard-Jones potentials. The Reactive Interface Force Field (IFF-R) is introduced, which allows the simulation of bond breaking using a substitution of harmonic bond stretching potentials with Morse bond potentials, maintains the high accuracy and speed of IFF, and eliminates discontinuities in bond energy upon dissociation. IFF-R also allows continuous simulations where bonds can break and form without needing a complex bond-order function, for example, using a combination with template-based reaction frameworks such as REACTER in LAMMPS. IFF-R is applicable to metals, ceramics, composites, and biopolymers such as cellulose and proteins. MD simulations with IFF-R were further used to create a dataset of over 1000 structures of pristine and defective CNT bundles and graphitic morphologies up to 80,000 atoms, as well as the associated stress-strain curves up to failure. The data set was used to develop and test machine-learning (ML) algorithms to predict the tensile modulus and strength for CNT assemblies, CNT junctions, and carbon-fiber cross-sections from the 3D atomic structure in up to about 1,000 faster speed than MD simulations. The most reliable ML method, called the hierarchical spatial graph neural network (HS-GNN), contains spatial information on bonds and torsions, as well as multi-scale resolution to include long-range interactions. HS-GNN performs well for sparse data and topologies outside the training range. The ML methods can be used to quickly screen computer generated 3D graphitic models for best mechanical properties to inform choices for composite materials, and can be extended to other chemistries and nanostructures.
■590 ▼aSchool code: 0051.
■650 4▼aComputational chemistry
■650 4▼aMaterials science
■653 ▼aCarbon
■653 ▼aComposites
■653 ▼aForce field
■653 ▼aMachine learning
■653 ▼aMolecular dynamics
■653 ▼aInterface force field
■690 ▼a0219
■690 ▼a0794
■690 ▼a0800
■71020▼aUniversity of Colorado at Boulder▼bMaterials Science and Engineering.
■7730 ▼tDissertations Abstracts International▼g85-11B.
■790 ▼a0051
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17160974▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
