서브메뉴
검색
Data-Driven Geometric Field Prediction Methods and Engineering Applications
Data-Driven Geometric Field Prediction Methods and Engineering Applications
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202103506
- ISBN
- 9798288853098
- DDC
- 621
- 서명/저자
- Data-Driven Geometric Field Prediction Methods and Engineering Applications
- 발행사항
- [Sl] : Carnegie Mellon University, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 124 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-01, Section: B.
- 주기사항
- Advisor: Kara, Levent.
- 학위논문주기
- Thesis (Ph.D.)--Carnegie Mellon University, 2025.
- 초록/해제
- 요약Design is a familiar iterative process across engineering disciplines. A well-informed design or optimization cycle guides an engineer from rough initial sketches and back-of-the-envelope calculations to polished final products accompanied by thorough, quantitative evaluations. However, the analyses that must be performed to assess a proposed design often involve computationally expensive computer simulations. Particularly in the case of part design, these simulations can take on the order of hours or days. An engineer in the early stages of a design cycle would benefit from a way to more quickly generate simulation results to expedite lateral design exploration.In this thesis, we propose machine learning methods to predict output fields on arbitrary input geometries. For example, given a 2-D or 3-D part, our methods can estimate stress or temperature throughout the part nearly instantaneously, without having to wait for a simulation to complete. The main draw of our methods is that they allow truly arbitrary input structure, i.e. the part need not be constrained to a uniform pixel or voxel grid space. Instead, any mesh-like geometric representation can serve as an input shape, making our models especially suitable in typical shape design settings.First, we propose a method inspired by image segmentation methods, but incorporating differentiable interpolation steps at multiple resolutions. We demonstrate this method on a von Mises stress field prediction problem for a 2-D part undergoing compression, and we show that the results achieved are superior to a U-Net architecture of similar capacity.Next, a Topology-Agnostic Graph U-Net method is defined, which we call TAG U-Net. This model uses graph convolution rather than image convolution, making it more flexible with respect to structure of input data. TAG U-Net is better suited to 3-D problems, and its performance is used to predict 3-D laser powder bed fusion simulation results, where it outperforms a standard graph neural network. Developing upon TAG U-Net, we extend our methods to take as input material properties in addition to part geometry. We show that by fine-tuning a pre-trained model, only a few new simulations are required to adapt the model to a new material context.Finally, we investigate materials science applications of geometric field prediction methods, demonstrating how field prediction can be used in denoising diffusion probabilistic models, specifically to predict material properties by learning from molecular simulation data.This thesis offers a look at the predictive capabilities of several models whose input data are geometric objects without consistent structure. By demonstrating how our work can be used for solid mechanics, additive manufacturing, and materials science, we emphasize the cross-disciplinary value of the proposed methods. The datasets used have been made publicly accessible to maximize their utility to the ML and design communities.
- 일반주제명
- Mechanical engineering
- 일반주제명
- Materials science
- 키워드
- Field prediction
- 기타저자
- Carnegie Mellon University Mechanical Engineering
- 기본자료저록
- Dissertations Abstracts International. 87-01B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017357401
■00520260202103506
■006m o d
■007cr#unu||||||||
■020 ▼a9798288853098
■035 ▼a(MiAaPQ)AAI32002790
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a621
■1001 ▼aFerguson, Kevin M.▼0(orcid)0009-0004-2234-4207
■24510▼aData-Driven Geometric Field Prediction Methods and Engineering Applications
■260 ▼a[Sl]▼bCarnegie Mellon University▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a124 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-01, Section: B.
■500 ▼aAdvisor: Kara, Levent.
■5021 ▼aThesis (Ph.D.)--Carnegie Mellon University, 2025.
■520 ▼aDesign is a familiar iterative process across engineering disciplines. A well-informed design or optimization cycle guides an engineer from rough initial sketches and back-of-the-envelope calculations to polished final products accompanied by thorough, quantitative evaluations. However, the analyses that must be performed to assess a proposed design often involve computationally expensive computer simulations. Particularly in the case of part design, these simulations can take on the order of hours or days. An engineer in the early stages of a design cycle would benefit from a way to more quickly generate simulation results to expedite lateral design exploration.In this thesis, we propose machine learning methods to predict output fields on arbitrary input geometries. For example, given a 2-D or 3-D part, our methods can estimate stress or temperature throughout the part nearly instantaneously, without having to wait for a simulation to complete. The main draw of our methods is that they allow truly arbitrary input structure, i.e. the part need not be constrained to a uniform pixel or voxel grid space. Instead, any mesh-like geometric representation can serve as an input shape, making our models especially suitable in typical shape design settings.First, we propose a method inspired by image segmentation methods, but incorporating differentiable interpolation steps at multiple resolutions. We demonstrate this method on a von Mises stress field prediction problem for a 2-D part undergoing compression, and we show that the results achieved are superior to a U-Net architecture of similar capacity.Next, a Topology-Agnostic Graph U-Net method is defined, which we call TAG U-Net. This model uses graph convolution rather than image convolution, making it more flexible with respect to structure of input data. TAG U-Net is better suited to 3-D problems, and its performance is used to predict 3-D laser powder bed fusion simulation results, where it outperforms a standard graph neural network. Developing upon TAG U-Net, we extend our methods to take as input material properties in addition to part geometry. We show that by fine-tuning a pre-trained model, only a few new simulations are required to adapt the model to a new material context.Finally, we investigate materials science applications of geometric field prediction methods, demonstrating how field prediction can be used in denoising diffusion probabilistic models, specifically to predict material properties by learning from molecular simulation data.This thesis offers a look at the predictive capabilities of several models whose input data are geometric objects without consistent structure. By demonstrating how our work can be used for solid mechanics, additive manufacturing, and materials science, we emphasize the cross-disciplinary value of the proposed methods. The datasets used have been made publicly accessible to maximize their utility to the ML and design communities.
■590 ▼aSchool code: 0041.
■650 4▼aMechanical engineering
■650 4▼aMaterials science
■653 ▼aAdditive manufacturing
■653 ▼aField prediction
■653 ▼aGeometric deep learning
■653 ▼aGraph neural networks
■653 ▼aSurrogate modeling
■690 ▼a0548
■690 ▼a0800
■690 ▼a0794
■71020▼aCarnegie Mellon University▼bMechanical Engineering.
■7730 ▼tDissertations Abstracts International▼g87-01B.
■790 ▼a0041
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17357401▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
