서브메뉴
검색
Dataset Construction and Characterization for Learning Automotive Aerodynamics
Dataset Construction and Characterization for Learning Automotive Aerodynamics
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202104740
- ISBN
- 9798290651910
- DDC
- 629.222
- 저자명
- Benjamin, Mark.
- 서명/저자
- Dataset Construction and Characterization for Learning Automotive Aerodynamics
- 발행사항
- [Sl] : Stanford University, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 121 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-03, Section: B.
- 주기사항
- Advisor: Iaccarino, Gianluca.
- 학위논문주기
- Thesis (Ph.D.)--Stanford University, 2024.
- 초록/해제
- 요약The cost of numerical simulations for automotive aerodynamics optimization makes data-driven surrogates an appealing alternative. However, there is a data insufficiency problem in this application space. Owing to the proprietary nature of commercial automotive designs, training datasets for drag prediction models are limited to a few freely-available, realistic geometries. Additionally, existing datasets suffer from a lack of quantification of intrinsic properties-such as diversity-that correlate with predictive performance on unseen samples; in other words, there is a dataset characterization problem.In this work, a database construction strategy is proposed that allows for the controlled generation of an arbitrary number of samples, by convex interpolation between a small number of basis geometries. A set of geometries is then constructed using this method, using the DrivAer automotive reference geometry for the basis cases. Large-eddy simulations are used to obtain the aerodynamic quantities of interest for each geometry. Convolutional neural network-based models are trained with this dataset to make aerodynamics predictions. The proposed method allows for a characterization of datasets based on size, density, and diversity. A formal measure of diversity for a given dataset is developed. Then, datasets of successively increasing diversity but constant size are constructed, and it is shown that the dataset diversity has an impact on the predictive accuracy of the neural networks. The proposed method allows for more rigorous a priori evaluation of models than is currently possible and can be applied readily to other problems of shape optimization.
- 일반주제명
- Automobiles
- 일반주제명
- Design optimization
- 일반주제명
- Multidimensional scaling
- 일반주제명
- Industrial research
- 일반주제명
- Fluid dynamics
- 일반주제명
- Reynolds number
- 일반주제명
- Energy consumption
- 일반주제명
- Geometry
- 일반주제명
- Aerodynamics
- 일반주제명
- Neural networks
- 일반주제명
- Aeronomy
- 일반주제명
- Mechanical engineering
- 기타저자
- Stanford University.
- 기본자료저록
- Dissertations Abstracts International. 87-03B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2024 us c eng d■001000017358709
■00520260202104740
■006m o d
■007cr#unu||||||||
■020 ▼a9798290651910
■035 ▼a(MiAaPQ)AAI32149699
■035 ▼a(MiAaPQ)Stanfordnx392gr8211
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a629.222
■1001 ▼aBenjamin, Mark.
■24510▼aDataset Construction and Characterization for Learning Automotive Aerodynamics
■260 ▼a[Sl]▼bStanford University▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a121 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-03, Section: B.
■500 ▼aAdvisor: Iaccarino, Gianluca.
■5021 ▼aThesis (Ph.D.)--Stanford University, 2024.
■520 ▼aThe cost of numerical simulations for automotive aerodynamics optimization makes data-driven surrogates an appealing alternative. However, there is a data insufficiency problem in this application space. Owing to the proprietary nature of commercial automotive designs, training datasets for drag prediction models are limited to a few freely-available, realistic geometries. Additionally, existing datasets suffer from a lack of quantification of intrinsic properties-such as diversity-that correlate with predictive performance on unseen samples; in other words, there is a dataset characterization problem.In this work, a database construction strategy is proposed that allows for the controlled generation of an arbitrary number of samples, by convex interpolation between a small number of basis geometries. A set of geometries is then constructed using this method, using the DrivAer automotive reference geometry for the basis cases. Large-eddy simulations are used to obtain the aerodynamic quantities of interest for each geometry. Convolutional neural network-based models are trained with this dataset to make aerodynamics predictions. The proposed method allows for a characterization of datasets based on size, density, and diversity. A formal measure of diversity for a given dataset is developed. Then, datasets of successively increasing diversity but constant size are constructed, and it is shown that the dataset diversity has an impact on the predictive accuracy of the neural networks. The proposed method allows for more rigorous a priori evaluation of models than is currently possible and can be applied readily to other problems of shape optimization.
■590 ▼aSchool code: 0212.
■650 4▼aAutomobiles
■650 4▼aDesign optimization
■650 4▼aMultidimensional scaling
■650 4▼aIndustrial research
■650 4▼aFluid dynamics
■650 4▼aReynolds number
■650 4▼aEnergy consumption
■650 4▼aGeometry
■650 4▼aAerodynamics
■650 4▼aNeural networks
■650 4▼aAeronomy
■650 4▼aMechanical engineering
■653 ▼aAutomotive aerodynamics
■653 ▼aDataset construction
■690 ▼a0367
■690 ▼a0548
■71020▼aStanford University.
■7730 ▼tDissertations Abstracts International▼g87-03B.
■790 ▼a0212
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17358709▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


