서브메뉴
검색
Machine Learning Methods to Optimize the Geometry and Topology of Meshes
Machine Learning Methods to Optimize the Geometry and Topology of Meshes
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20250211151435
- ISBN
- 9798384447771
- DDC
- 621
- 서명/저자
- Machine Learning Methods to Optimize the Geometry and Topology of Meshes
- 발행사항
- [Sl] : University of California, Berkeley, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 125 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-03, Section: B.
- 주기사항
- Advisor: Persson, Per-Olof;Shadden, Shawn.
- 학위논문주기
- Thesis (Ph.D.)--University of California, Berkeley, 2024.
- 초록/해제
- 요약Meshes are used ubiquitously in engineering for representing geometries, performing computational simulations, and generating computer graphics renderings. Automatically generating suitable meshes for downstream applications remains a key bottleneck in many workflows and often requires significant manual intervention. It is challenging to optimize mesh data-structures because they can be highly unstructured in their most general form.A mesh has two fundamental attributes - geometry and topology. Geometry deals with the position and shape of objects in space. Topology is concerned with the connectivity of mesh elements. It is essential to optimize both of these attributes to generate a desirable mesh for a target application such as simulation. This dissertation explores machine learning methods to optimize both of these attributes.The first part of this dissertation is concerned with mesh topology. We will describe a deep reinforcement learning framework to optimize the topology of 2D meshes using elementary mesh editing operations. The framework is trained purely in self-play reinforcement learning to optimize a given user defined objective function. We describe a novel neural network architecture that is able to encode the local topology of a mesh around a given mesh neighborhood. Subsequently, the neural network is trained to predict a probability distribution over the local action space in order to maximize the cumulative reward as prescribed by the given objective function. The agent is trained on randomly generated 2D polygonal shapes. We demonstrate generalization to inputs that were never seen during training. The proposed framework is particularly effective at coarse block decomposition of polygonal shapes where the aim is to minimize the number of irregular vertices in the mesh.We will then tackle the problem of geometry. We describe a deep learning method to automatically generate patient-specific, simulation ready 3D surface meshes of the human heart directly from clinical imaging. The proposed method is a two-stage mesh deformation process that transforms a given template mesh to match the underlying target geometry in the image data. The first stage consists of a learned affine transformation conditioned on the input image. This stage is trained to roughly align the template in terms of scale and orientation to the image data. The second stage consists of a learned local diffeomorphic deformation field conditioned on the image and the current location of the template. This stage improves the accuracy of the prediction by capturing finer details of the target geometry. We describe a novel loss function derived from the kinematics of motion of continuous bodies that penalizes undesirable phenomenon such as surface interpenetration resulting in anatomically accurate, physically realistic, simulation ready meshes. The proposed framework is validated against a large held-out test dataset and compared with prior state-of-the-art along a variety of accuracy and quality metrics.
- 일반주제명
- Mechanical engineering
- 일반주제명
- Computer science
- 일반주제명
- Applied mathematics
- 키워드
- Machine learning
- 키워드
- Mesh generation
- 키워드
- Mesh topology
- 기타저자
- University of California, Berkeley Mechanical Engineering
- 기본자료저록
- Dissertations Abstracts International. 86-03B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008250123s2024 us c eng d■001000017161715
■00520250211151435
■006m o d
■007cr#unu||||||||
■020 ▼a9798384447771
■035 ▼a(MiAaPQ)AAI31295553
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a621
■1001 ▼aNarayanan, Arjun.
■24510▼aMachine Learning Methods to Optimize the Geometry and Topology of Meshes
■260 ▼a[Sl]▼bUniversity of California, Berkeley▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a125 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-03, Section: B.
■500 ▼aAdvisor: Persson, Per-Olof;Shadden, Shawn.
■5021 ▼aThesis (Ph.D.)--University of California, Berkeley, 2024.
■520 ▼aMeshes are used ubiquitously in engineering for representing geometries, performing computational simulations, and generating computer graphics renderings. Automatically generating suitable meshes for downstream applications remains a key bottleneck in many workflows and often requires significant manual intervention. It is challenging to optimize mesh data-structures because they can be highly unstructured in their most general form.A mesh has two fundamental attributes - geometry and topology. Geometry deals with the position and shape of objects in space. Topology is concerned with the connectivity of mesh elements. It is essential to optimize both of these attributes to generate a desirable mesh for a target application such as simulation. This dissertation explores machine learning methods to optimize both of these attributes.The first part of this dissertation is concerned with mesh topology. We will describe a deep reinforcement learning framework to optimize the topology of 2D meshes using elementary mesh editing operations. The framework is trained purely in self-play reinforcement learning to optimize a given user defined objective function. We describe a novel neural network architecture that is able to encode the local topology of a mesh around a given mesh neighborhood. Subsequently, the neural network is trained to predict a probability distribution over the local action space in order to maximize the cumulative reward as prescribed by the given objective function. The agent is trained on randomly generated 2D polygonal shapes. We demonstrate generalization to inputs that were never seen during training. The proposed framework is particularly effective at coarse block decomposition of polygonal shapes where the aim is to minimize the number of irregular vertices in the mesh.We will then tackle the problem of geometry. We describe a deep learning method to automatically generate patient-specific, simulation ready 3D surface meshes of the human heart directly from clinical imaging. The proposed method is a two-stage mesh deformation process that transforms a given template mesh to match the underlying target geometry in the image data. The first stage consists of a learned affine transformation conditioned on the input image. This stage is trained to roughly align the template in terms of scale and orientation to the image data. The second stage consists of a learned local diffeomorphic deformation field conditioned on the image and the current location of the template. This stage improves the accuracy of the prediction by capturing finer details of the target geometry. We describe a novel loss function derived from the kinematics of motion of continuous bodies that penalizes undesirable phenomenon such as surface interpenetration resulting in anatomically accurate, physically realistic, simulation ready meshes. The proposed framework is validated against a large held-out test dataset and compared with prior state-of-the-art along a variety of accuracy and quality metrics.
■590 ▼aSchool code: 0028.
■650 4▼aMechanical engineering
■650 4▼aComputer science
■650 4▼aApplied mathematics
■653 ▼aMachine learning
■653 ▼aMesh generation
■653 ▼aReinforcement learning
■653 ▼aMesh topology
■690 ▼a0548
■690 ▼a0984
■690 ▼a0364
■71020▼aUniversity of California, Berkeley▼bMechanical Engineering.
■7730 ▼tDissertations Abstracts International▼g86-03B.
■790 ▼a0028
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17161715▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


