서브메뉴
검색
Application-Driven Design of Machine Learning Surrogate Models in Catalysis
Application-Driven Design of Machine Learning Surrogate Models in Catalysis
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202105304
- ISBN
- 9798270247058
- DDC
- 660
- 서명/저자
- Application-Driven Design of Machine Learning Surrogate Models in Catalysis
- 발행사항
- [Sl] : Carnegie Mellon University, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 96 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-06, Section: B.
- 주기사항
- Advisor: Kitchin, John R.;Laird, Carl D.
- 학위논문주기
- Thesis (Ph.D.)--Carnegie Mellon University, 2025.
- 초록/해제
- 요약The latest developments in computing and data science have enabled the use of highly parameterized, general-purpose regressors called machine learning models. While they are useful "universal approximators" out of the box, their architectures can also be tailored for specific applications. The computational catalysis community has embraced these machine learning models as fast surrogates for Density Functional Theory (DFT), a physics-based atomistic simulation technique which is fairly accurate but very computationally intensive. These surrogates are now competitively accurate with DFT and offer an impressive increase in speed.However, much work remains to be done connecting these atomistic simulations to other models based on physics, process design, or economics. Fortunately, the use of surrogates actually gives us an opportunity to control the functional form of the model and design one with desired properties. In this work, I will present three examples of this, in which a surrogate is carefully designed for a specific engineering application. First, a graph neural network is modified to provide physically realistic predictions of electron density, which is necessary to use this quantity for further quantitative analysis. Second, piecewise linear surrogates are used to connect disparate models into a unified framework that admits cross-scale optimization. Finally, I introduce a new type of piecewise linear model which is flexible, accurate, and fast compared to other existing methods. The surrogate approaches represent opportunities to tailor computational models for specific applications of interest.
- 일반주제명
- Chemical engineering
- 일반주제명
- Computational chemistry
- 일반주제명
- Computational physics
- 키워드
- Decision trees
- 키워드
- Electron density
- 키워드
- Hyperplanes
- 키워드
- Optimization
- 키워드
- Surrogate models
- 기타저자
- Carnegie Mellon University Chemical Engineering
- 기본자료저록
- Dissertations Abstracts International. 87-06B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017360101
■00520260202105304
■006m o d
■007cr#unu||||||||
■020 ▼a9798270247058
■035 ▼a(MiAaPQ)AAI32282596
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a660
■1001 ▼aSunshine, Ethan M.
■24510▼aApplication-Driven Design of Machine Learning Surrogate Models in Catalysis
■260 ▼a[Sl]▼bCarnegie Mellon University▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a96 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-06, Section: B.
■500 ▼aAdvisor: Kitchin, John R.;Laird, Carl D.
■5021 ▼aThesis (Ph.D.)--Carnegie Mellon University, 2025.
■520 ▼aThe latest developments in computing and data science have enabled the use of highly parameterized, general-purpose regressors called machine learning models. While they are useful "universal approximators" out of the box, their architectures can also be tailored for specific applications. The computational catalysis community has embraced these machine learning models as fast surrogates for Density Functional Theory (DFT), a physics-based atomistic simulation technique which is fairly accurate but very computationally intensive. These surrogates are now competitively accurate with DFT and offer an impressive increase in speed.However, much work remains to be done connecting these atomistic simulations to other models based on physics, process design, or economics. Fortunately, the use of surrogates actually gives us an opportunity to control the functional form of the model and design one with desired properties. In this work, I will present three examples of this, in which a surrogate is carefully designed for a specific engineering application. First, a graph neural network is modified to provide physically realistic predictions of electron density, which is necessary to use this quantity for further quantitative analysis. Second, piecewise linear surrogates are used to connect disparate models into a unified framework that admits cross-scale optimization. Finally, I introduce a new type of piecewise linear model which is flexible, accurate, and fast compared to other existing methods. The surrogate approaches represent opportunities to tailor computational models for specific applications of interest.
■590 ▼aSchool code: 0041.
■650 4▼aChemical engineering
■650 4▼aComputational chemistry
■650 4▼aComputational physics
■653 ▼aDecision trees
■653 ▼aElectron density
■653 ▼aGraph neural networks
■653 ▼aHyperplanes
■653 ▼aOptimization
■653 ▼aSurrogate models
■690 ▼a0542
■690 ▼a0219
■690 ▼a0216
■71020▼aCarnegie Mellon University▼bChemical Engineering.
■7730 ▼tDissertations Abstracts International▼g87-06B.
■790 ▼a0041
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17360101▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
