서브메뉴
검색
Bridging Text, Graph, and Action: AI-Assisted Catalyst Discovery
Bridging Text, Graph, and Action: AI-Assisted Catalyst Discovery
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202104812
- ISBN
- 9798291556160
- DDC
- 660
- 저자명
- Ock, Janghoon.
- 서명/저자
- Bridging Text, Graph, and Action: AI-Assisted Catalyst Discovery
- 발행사항
- [Sl] : Carnegie Mellon University, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 154 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-02, Section: B.
- 주기사항
- Advisor: Barati Farimani, Amir.
- 학위논문주기
- Thesis (Ph.D.)--Carnegie Mellon University, 2025.
- 초록/해제
- 요약Catalysts enable more efficient energy and chemical transformations by lowering reaction barriers. Identifying the optimal catalyst for a given application is essential to reduce energy consumption and improve process profitability. Recently, machine learning (ML) has made remarkable advances in computational catalysis, particularly through the development of graph neural networks (GNNs) that approximate quantum chemistry methods such as density functional theory (DFT). Although much of the focus has been on improving the accuracy of the model in predicting energy and force, it is equally important to ensure the practical applicability of these models, expand the types of usable input data, and reduce the need for manual tool implementation through greater automation. This thesis develops ML frameworks to advance catalyst discovery in three directions. First, we introduce a Transformer-based language model that predicts the adsorption energy from textual inputs, offering a potential structure-agnostic alternative. The model can process various forms of text, including formatted strings and natural language descriptions, while achieving a competitive level of accuracy. We further enhance its accuracy by aligning its latent space with that of GNNs through graph-assisted pre-training, reducing the mean absolute error by up to 9.8%. This strategy presents an approach for integrating multiple catalyst modalities in a shared representation space. Finally, we develop a large language model (LLM)-powered agent that autonomously identifies stable adsorption configurations using built-in chemical knowledge and reasoning capabilities. This agent substantially reduces the number of configurations required to locate global energy minima-utilizing, on average, only 27% of the initial candidate set compared to algorithmic baselines. Together, these contributions advance AI-assisted catalyst discovery by enabling faster, more interpretable, and less labor-intensive workflows.
- 일반주제명
- Chemical engineering
- 일반주제명
- Energy
- 일반주제명
- Alternative energy
- 키워드
- Machine learning
- 기타저자
- Carnegie Mellon University Chemical Engineering
- 기본자료저록
- Dissertations Abstracts International. 87-02B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017358937
■00520260202104812
■006m o d
■007cr#unu||||||||
■020 ▼a9798291556160
■035 ▼a(MiAaPQ)AAI32167434
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a660
■1001 ▼aOck, Janghoon.▼0(orcid)0009-0000-0370-4212
■24510▼aBridging Text, Graph, and Action: AI-Assisted Catalyst Discovery
■260 ▼a[Sl]▼bCarnegie Mellon University▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a154 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-02, Section: B.
■500 ▼aAdvisor: Barati Farimani, Amir.
■5021 ▼aThesis (Ph.D.)--Carnegie Mellon University, 2025.
■520 ▼aCatalysts enable more efficient energy and chemical transformations by lowering reaction barriers. Identifying the optimal catalyst for a given application is essential to reduce energy consumption and improve process profitability. Recently, machine learning (ML) has made remarkable advances in computational catalysis, particularly through the development of graph neural networks (GNNs) that approximate quantum chemistry methods such as density functional theory (DFT). Although much of the focus has been on improving the accuracy of the model in predicting energy and force, it is equally important to ensure the practical applicability of these models, expand the types of usable input data, and reduce the need for manual tool implementation through greater automation. This thesis develops ML frameworks to advance catalyst discovery in three directions. First, we introduce a Transformer-based language model that predicts the adsorption energy from textual inputs, offering a potential structure-agnostic alternative. The model can process various forms of text, including formatted strings and natural language descriptions, while achieving a competitive level of accuracy. We further enhance its accuracy by aligning its latent space with that of GNNs through graph-assisted pre-training, reducing the mean absolute error by up to 9.8%. This strategy presents an approach for integrating multiple catalyst modalities in a shared representation space. Finally, we develop a large language model (LLM)-powered agent that autonomously identifies stable adsorption configurations using built-in chemical knowledge and reasoning capabilities. This agent substantially reduces the number of configurations required to locate global energy minima-utilizing, on average, only 27% of the initial candidate set compared to algorithmic baselines. Together, these contributions advance AI-assisted catalyst discovery by enabling faster, more interpretable, and less labor-intensive workflows.
■590 ▼aSchool code: 0041.
■650 4▼aChemical engineering
■650 4▼aEnergy
■650 4▼aAlternative energy
■653 ▼aMachine learning
■653 ▼aGraph neural networks
■653 ▼aLarge language model
■653 ▼aDensity functional theory
■653 ▼aEnergy consumption
■690 ▼a0542
■690 ▼a0800
■690 ▼a0363
■690 ▼a0791
■71020▼aCarnegie Mellon University▼bChemical Engineering.
■7730 ▼tDissertations Abstracts International▼g87-02B.
■790 ▼a0041
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17358937▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
