서브메뉴
검색
Physics-Constrained Machine Learning for Scale-Bridging in Crystalline Materials
Physics-Constrained Machine Learning for Scale-Bridging in Crystalline Materials
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202105236
- ISBN
- 9798291567814
- DDC
- 530
- 저자명
- Holber, Jamie.
- 서명/저자
- Physics-Constrained Machine Learning for Scale-Bridging in Crystalline Materials
- 발행사항
- [Sl] : University of Michigan, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 162 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-03, Section: B.
- 주기사항
- Advisor: Garikipati, Krishna.
- 학위논문주기
- Thesis (Ph.D.)--University of Michigan, 2025.
- 초록/해제
- 요약Understanding critical phenomena in materials, such as mechano-chemical phase transitions and the emergence of multiple phases or variants, is essential for predicting material behavior under real-world conditions. This is particularly important for modeling lithium-ion battery components, such as crystalline cathodes, which undergo continuous phase transitions due to lithium flux during charging and discharging. Phase transitions influence battery capacity, degradation, and overall performance. Understanding the underlying mechanisms is crucial for designing and utilizing lithium batteries, which are a vital part of our energy infrastructure. Phase transitions are driven by the free energy of the system. Traditionally, free energy functions have been phenomenologically derived and fitted to macroscopic behavior, limiting their ability to capture atomistic effects. Moreover, most computational material methods focus on a single scale: density functional theory (DFT) models the electronic scale; Monte Carlo simulations describe the atomistic scale; and phase field methods describe continuum scale behavior. Data-driven machine learning (ML) methods provide a powerful pathway for bridging scales. We have developed a machine learning enabled scale-bridging framework to incorporate atomistic energy into high-dimensional free energy representations. The free energy can then be used to inform phase evolution in continuum scale modeling. Our methodology begins with DFT calculations to characterize the formation energies at the electronic scale. Then ML methods such as cluster expansions or graph neural networks (GNNs) are trained with DFT data to predict formation energies for an arbitrary atomistic configuration. These formation energy models enable efficient Monte Carlo simulations to sample relevant atomic configurations. The resulting simulations provide a direct relationship between order parameters and free energy derivatives. Next, integrable deep neural networks (IDNNs) or integrable graph neural networks (IGNNs) are trained to the Monte Carlo data to yield high-dimensional free energy surfaces dependent on phase orderings and lithium compositions. These free energy representations are used to drive the phase-field simulations. This dissertation presents several improvements to the scale-bridging framework, including a set of enhanced active learning strategies for sampling the high-dimensional Monte Carlo data and the introduction of GNNs for modeling formation energies and free energies for increased flexibility in representing a diverse set of materials and more complex ordering phenomena.
- 일반주제명
- Physics
- 일반주제명
- Materials science
- 일반주제명
- Condensed matter physics
- 일반주제명
- Computational physics
- 키워드
- Scale bridging
- 키워드
- Active learning
- 기타저자
- University of Michigan Applied Physics
- 기본자료저록
- Dissertations Abstracts International. 87-03B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017359920
■00520260202105236
■006m o d
■007cr#unu||||||||
■020 ▼a9798291567814
■035 ▼a(MiAaPQ)AAI32271951
■035 ▼a(MiAaPQ)umichrackham006319
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a530
■1001 ▼aHolber, Jamie.
■24510▼aPhysics-Constrained Machine Learning for Scale-Bridging in Crystalline Materials
■260 ▼a[Sl]▼bUniversity of Michigan▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a162 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-03, Section: B.
■500 ▼aAdvisor: Garikipati, Krishna.
■5021 ▼aThesis (Ph.D.)--University of Michigan, 2025.
■520 ▼aUnderstanding critical phenomena in materials, such as mechano-chemical phase transitions and the emergence of multiple phases or variants, is essential for predicting material behavior under real-world conditions. This is particularly important for modeling lithium-ion battery components, such as crystalline cathodes, which undergo continuous phase transitions due to lithium flux during charging and discharging. Phase transitions influence battery capacity, degradation, and overall performance. Understanding the underlying mechanisms is crucial for designing and utilizing lithium batteries, which are a vital part of our energy infrastructure. Phase transitions are driven by the free energy of the system. Traditionally, free energy functions have been phenomenologically derived and fitted to macroscopic behavior, limiting their ability to capture atomistic effects. Moreover, most computational material methods focus on a single scale: density functional theory (DFT) models the electronic scale; Monte Carlo simulations describe the atomistic scale; and phase field methods describe continuum scale behavior. Data-driven machine learning (ML) methods provide a powerful pathway for bridging scales. We have developed a machine learning enabled scale-bridging framework to incorporate atomistic energy into high-dimensional free energy representations. The free energy can then be used to inform phase evolution in continuum scale modeling. Our methodology begins with DFT calculations to characterize the formation energies at the electronic scale. Then ML methods such as cluster expansions or graph neural networks (GNNs) are trained with DFT data to predict formation energies for an arbitrary atomistic configuration. These formation energy models enable efficient Monte Carlo simulations to sample relevant atomic configurations. The resulting simulations provide a direct relationship between order parameters and free energy derivatives. Next, integrable deep neural networks (IDNNs) or integrable graph neural networks (IGNNs) are trained to the Monte Carlo data to yield high-dimensional free energy surfaces dependent on phase orderings and lithium compositions. These free energy representations are used to drive the phase-field simulations. This dissertation presents several improvements to the scale-bridging framework, including a set of enhanced active learning strategies for sampling the high-dimensional Monte Carlo data and the introduction of GNNs for modeling formation energies and free energies for increased flexibility in representing a diverse set of materials and more complex ordering phenomena.
■590 ▼aSchool code: 0127.
■650 4▼aPhysics
■650 4▼aMaterials science
■650 4▼aCondensed matter physics
■650 4▼aComputational physics
■653 ▼aPhysics-informed machine learning
■653 ▼aScale bridging
■653 ▼aActive learning
■653 ▼aGraph neural networks
■653 ▼aDensity functional theory
■690 ▼a0605
■690 ▼a0794
■690 ▼a0800
■690 ▼a0611
■690 ▼a0216
■71020▼aUniversity of Michigan▼bApplied Physics.
■7730 ▼tDissertations Abstracts International▼g87-03B.
■790 ▼a0127
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17359920▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
