서브메뉴
검색
Overcoming Small Datasets in Machine Learning Studies of Multi-Physics Flows in Propulsion.
Overcoming Small Datasets in Machine Learning Studies of Multi-Physics Flows in Propulsion.
상세정보
- 자료유형
- 학위논문
- Control Number
- 0017162986
- International Standard Book Number
- 9798384337959
- Dewey Decimal Classification Number
- 530
- Main Entry-Personal Name
- Chung, Wai Tong.
- Publication, Distribution, etc. (Imprint
- [S.l.] : Stanford University., 2024
- Publication, Distribution, etc. (Imprint
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- Physical Description
- 182 p.
- General Note
- Source: Dissertations Abstracts International, Volume: 86-03, Section: B.
- General Note
- Advisor: Ihme, Matthias.
- Dissertation Note
- Thesis (Ph.D.)--Stanford University, 2024.
- Summary, Etc.
- 요약While machine learning (ML) methods can offer numerous opportunities in modeling multi-physics flows, these approaches often rely on the availability of large datasets for generating reliable predictions. This can be challenging for propulsion applications, especially since data generated by industrial sensors, experiments, and numerical simulations of flow phenomena in propulsion systems can be challenging to collect.In this dissertation, we directly address current gaps in data availability by developing a 2.2 TB ML dataset from 34 high-fidelity direct numerical simulations (DNS) of turbulent flows. We employ this data for benchmarking super-resolution of turbulent flows, and provide insights into the role of different deep learning designs and computational scale in a popular ML application within multi-physics flows.To address issues in accessing data from relatively under-explored flow configurations (e.g., real, hypersonic, and multiphase fluids) in propulsion systems, we investigate opportunities offered by linear regression and random forest models in modeling subgrid-scale (SGS) closure on small turbulent transcritical DNS dataset. Through a priorianalysis, interpretable metrics from random forest models, along with weights within linear regressors, are shown to assist in discovering analytical expressions for modeling SGS stresses and a closure term that arises from a real-fluid equation-of-state.To ameliorate spurious errors that can arise when integrating insufficiently trained ML models within multi-physics flow solvers, we develop a strategy involving an MLbased classifier that assigns three different combustion models of varying fidelity and cost within a shared simulation domain. Results from a posteriorisimulations show that this data-assisted framework demonstrates promise as a tool for controlling the fidelity-cost trade-off in numerical multi-physics flow simulations.Finally, we investigate the benefits of combining domain knowledge with ML by integrating a deep learning model with a stochastic differential equation for predicting the spatio-temporal behavior of laser ignition kernels with sparse ensemble data of a model rocket combustor. Results show that this hybrid reduced-order model can predict dominant ignition modes observed from corresponding experimental measurements, and generate spatially resolved ignition probability, at lower costs than high-fidelity turbulent reacting flow simulations approaches.Overall, the efforts within this dissertation contribute towards overcoming data limitations in ML-based modeling within science and engineering, specifically in the context of multi-physics flows found in propulsion systems.
- Subject Added Entry-Topical Term
- Physics.
- Subject Added Entry-Topical Term
- Viscosity.
- Subject Added Entry-Topical Term
- Mathematical models.
- Subject Added Entry-Topical Term
- Neural networks.
- Subject Added Entry-Topical Term
- Energy.
- Subject Added Entry-Topical Term
- Reynolds number.
- Subject Added Entry-Topical Term
- Fluid mechanics.
- Added Entry-Corporate Name
- Stanford University.
- Host Item Entry
- Dissertations Abstracts International. 86-03B.
- Electronic Location and Access
- 로그인을 한후 보실 수 있는 자료입니다.
- Control Number
- joongbu:655421
MARC
008250224s2024 us ||||||||||||||c||eng d■001000017162986
■00520250211152120
■006m o d
■007cr#unu||||||||
■020 ▼a9798384337959
■035 ▼a(MiAaPQ)AAI31460380
■035 ▼a(MiAaPQ)Stanfordxw487vj6602
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a530
■1001 ▼aChung, Wai Tong.
■24510▼aOvercoming Small Datasets in Machine Learning Studies of Multi-Physics Flows in Propulsion.
■260 ▼a[S.l.]▼bStanford University. ▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a182 p.
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-03, Section: B.
■500 ▼aAdvisor: Ihme, Matthias.
■5021 ▼aThesis (Ph.D.)--Stanford University, 2024.
■520 ▼aWhile machine learning (ML) methods can offer numerous opportunities in modeling multi-physics flows, these approaches often rely on the availability of large datasets for generating reliable predictions. This can be challenging for propulsion applications, especially since data generated by industrial sensors, experiments, and numerical simulations of flow phenomena in propulsion systems can be challenging to collect.In this dissertation, we directly address current gaps in data availability by developing a 2.2 TB ML dataset from 34 high-fidelity direct numerical simulations (DNS) of turbulent flows. We employ this data for benchmarking super-resolution of turbulent flows, and provide insights into the role of different deep learning designs and computational scale in a popular ML application within multi-physics flows.To address issues in accessing data from relatively under-explored flow configurations (e.g., real, hypersonic, and multiphase fluids) in propulsion systems, we investigate opportunities offered by linear regression and random forest models in modeling subgrid-scale (SGS) closure on small turbulent transcritical DNS dataset. Through a priorianalysis, interpretable metrics from random forest models, along with weights within linear regressors, are shown to assist in discovering analytical expressions for modeling SGS stresses and a closure term that arises from a real-fluid equation-of-state.To ameliorate spurious errors that can arise when integrating insufficiently trained ML models within multi-physics flow solvers, we develop a strategy involving an MLbased classifier that assigns three different combustion models of varying fidelity and cost within a shared simulation domain. Results from a posteriorisimulations show that this data-assisted framework demonstrates promise as a tool for controlling the fidelity-cost trade-off in numerical multi-physics flow simulations.Finally, we investigate the benefits of combining domain knowledge with ML by integrating a deep learning model with a stochastic differential equation for predicting the spatio-temporal behavior of laser ignition kernels with sparse ensemble data of a model rocket combustor. Results show that this hybrid reduced-order model can predict dominant ignition modes observed from corresponding experimental measurements, and generate spatially resolved ignition probability, at lower costs than high-fidelity turbulent reacting flow simulations approaches.Overall, the efforts within this dissertation contribute towards overcoming data limitations in ML-based modeling within science and engineering, specifically in the context of multi-physics flows found in propulsion systems.
■590 ▼aSchool code: 0212.
■650 4▼aPhysics.
■650 4▼aViscosity.
■650 4▼aMathematical models.
■650 4▼aNeural networks.
■650 4▼aEnergy.
■650 4▼aReynolds number.
■650 4▼aFluid mechanics.
■690 ▼a0791
■690 ▼a0605
■690 ▼a0800
■690 ▼a0204
■71020▼aStanford University.
■7730 ▼tDissertations Abstracts International▼g86-03B.
■790 ▼a0212
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17162986▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
미리보기
내보내기
chatGPT토론
Ai 추천 관련 도서
detalle info
- Reserva
- 캠퍼스간 도서대출
- 서가에 없는 책 신고
- Mi carpeta