서브메뉴
검색
Numerical Simulations and Machine Learning Surrogates for Predicting Solidification Microstructure in Additive Manufacturing
Numerical Simulations and Machine Learning Surrogates for Predicting Solidification Microstructure in Additive Manufacturing
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260311091549.5
- ISBN
- 9798270229054
- DDC
- 006
- 저자명
- Qin, Yigong
- 서명/저자
- Numerical Simulations and Machine Learning Surrogates for Predicting Solidification Microstructure in Additive Manufacturing / Yigong Qin
- 발행사항
- [Sl] : The University of Texas at Austin, 2025
- 형태사항
- 1 electronic resource (238 pages)
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-06, Section: B.
- 주기사항
- Advisors: Biros, George Committee members: Willcox, Karen; Ghattas, Omar; Seepersad, Carolyn.
- 학위논문주기
- - Ph.D. : The University of Texas at Austin, 2025.
- 초록/해제
- 요약Predicting microstructure evolution during alloy solidification is of great importance in additive manufacturing (AM) and other processes as the microstructure controls the mechanical properties of the built components. High fidelity full-order model simulations (FOMS) like phase field partial differential equations (PDEs) are an indispensable tool for process-to-mechanical-properties characterization. The inherent randomness of the initial grain structure necessitates ensemble simulations to predict microstructure statistics, such as grain size distributions. However, solving phase field equations can be computationally expensive as they require fine spatial and temporal discretizations. Their cost becomes an obstacle to parametric studies and ensemble runs, and ultimately makes downstream tasks like optimal control and uncertainty quantification challenging. This dissertation focuses on efficient surrogate models for predicting the solidified microstructure. Our first attempt is to reduce the cost of dendrite-resolved microstructure simulation in a 2D melt pool. We present an approximate "line" model that simulates the microstructure within a representative volume element of a much smaller size. We quantitatively assess the accuracy of the line model by comparing it with direct numerical simulations (DNS) at the full-melt-pool scale. We observe that for AM conditions the quantities of interest (QoIs) from the line model are in excellent agreement with the full DNS results for both shallow and deep melt pools. Our second approach is to use deep learning methods to accelerate the prediction of grain structures. We introduce GrainNN, a sequence-to-sequence long-short-term-memory (LSTM) network for 2D epitaxial grain growth. Its innovations include (1) hand-crafted grain features to represent 2D grains; (2) a self-attention mechanism for spatial grain interactions; and (3) algorithms to generalize to grain configurations and domain sizes that are unseen in training. GrainNN is orders of magnitude faster than phase field simulations and it is pointwise accurate with 5%-15% error. Further, we introduce GrainGNN, a graph neural network for 3D epitaxial grain growth. We use a graph with hand-crafted features to represent grain structure on the interface. GrainGNN uses a regressor to predict the grain boundary motion and a classifier to predict topological events. We train GrainGNN with phase field simulations and generalize it to a larger domain size with a larger number of grains. We investigate its error for different thermal parameters and initial grain configurations. GrainGNN is 150×-2000× faster than phase field simulations with 10%-20% error. Moreover, we extend GrainGNN to predict grain formation in moving melt pools by adding network features describing the melt pool geometry. We discuss its accuracy for different melt pool sizes, shapes, track lengths and we generalize it to multi-layer scans.
- 언어주기
- English
- 일반주제명
- Industrial engineering
- 일반주제명
- Materials science
- 일반주제명
- Mechanical engineering
- 키워드
- Microstructures
- 키워드
- Surrogate models
- 키워드
- GrainNN
- 기타저자
- The University of Texas at Austin Mechanical Engineering
- 기본자료저록
- Dissertations Abstracts International. 87-06B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260311s2025 us eng d■001000017361140
■00520260311091549.5
■006m o d
■007cr|nu||||||||
■020 ▼a9798270229054
■040 ▼aMiAaPQD▼beng▼cMiAaPQD▼erda
■082 ▼a006
■1001 ▼aQin, Yigong▼eauthor.
■24510▼aNumerical Simulations and Machine Learning Surrogates for Predicting Solidification Microstructure in Additive Manufacturing ▼cYigong Qin
■260 ▼a[Sl]▼bThe University of Texas at Austin▼c2025
■264 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a1 electronic resource (238 pages)
■336 ▼atext▼btxt▼2rdacontent
■337 ▼acomputer▼bc▼2rdamedia
■338 ▼aonline resource▼bcr▼2rdacarrier
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-06, Section: B.
■500 ▼aAdvisors: Biros, George Committee members: Willcox, Karen; Ghattas, Omar; Seepersad, Carolyn.
■5021 ▼bPh.D.▼cThe University of Texas at Austin▼d2025.
■520 ▼aPredicting microstructure evolution during alloy solidification is of great importance in additive manufacturing (AM) and other processes as the microstructure controls the mechanical properties of the built components. High fidelity full-order model simulations (FOMS) like phase field partial differential equations (PDEs) are an indispensable tool for process-to-mechanical-properties characterization. The inherent randomness of the initial grain structure necessitates ensemble simulations to predict microstructure statistics, such as grain size distributions. However, solving phase field equations can be computationally expensive as they require fine spatial and temporal discretizations. Their cost becomes an obstacle to parametric studies and ensemble runs, and ultimately makes downstream tasks like optimal control and uncertainty quantification challenging. This dissertation focuses on efficient surrogate models for predicting the solidified microstructure. Our first attempt is to reduce the cost of dendrite-resolved microstructure simulation in a 2D melt pool. We present an approximate "line" model that simulates the microstructure within a representative volume element of a much smaller size. We quantitatively assess the accuracy of the line model by comparing it with direct numerical simulations (DNS) at the full-melt-pool scale. We observe that for AM conditions the quantities of interest (QoIs) from the line model are in excellent agreement with the full DNS results for both shallow and deep melt pools. Our second approach is to use deep learning methods to accelerate the prediction of grain structures. We introduce GrainNN, a sequence-to-sequence long-short-term-memory (LSTM) network for 2D epitaxial grain growth. Its innovations include (1) hand-crafted grain features to represent 2D grains; (2) a self-attention mechanism for spatial grain interactions; and (3) algorithms to generalize to grain configurations and domain sizes that are unseen in training. GrainNN is orders of magnitude faster than phase field simulations and it is pointwise accurate with 5%-15% error. Further, we introduce GrainGNN, a graph neural network for 3D epitaxial grain growth. We use a graph with hand-crafted features to represent grain structure on the interface. GrainGNN uses a regressor to predict the grain boundary motion and a classifier to predict topological events. We train GrainGNN with phase field simulations and generalize it to a larger domain size with a larger number of grains. We investigate its error for different thermal parameters and initial grain configurations. GrainGNN is 150×-2000× faster than phase field simulations with 10%-20% error. Moreover, we extend GrainGNN to predict grain formation in moving melt pools by adding network features describing the melt pool geometry. We discuss its accuracy for different melt pool sizes, shapes, track lengths and we generalize it to multi-layer scans.
■546 ▼aEnglish
■590 ▼aSchool code: 0227
■650 4▼aIndustrial engineering
■650 4▼aMaterials science
■650 4▼aMechanical engineering
■653 ▼aAdditive manufacturing
■653 ▼aMicrostructures
■653 ▼aAlloy solidification
■653 ▼aSurrogate models
■653 ▼aGrainNN
■7102 ▼aThe University of Texas at Austin▼bMechanical Engineering.▼edegree granting institution.
■7201 ▼aBiros, George▼edegree supervisor.
■7730 ▼tDissertations Abstracts International▼g87-06B.
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17361140▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
