서브메뉴
검색
Towards Reliable AI for Materials Discovery at Scale
Towards Reliable AI for Materials Discovery at Scale
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202104836
- ISBN
- 9798297600669
- DDC
- 542
- 저자명
- Deng, Bowen.
- 서명/저자
- Towards Reliable AI for Materials Discovery at Scale
- 발행사항
- [Sl] : University of California, Berkeley, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 151 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-04, Section: B.
- 주기사항
- Advisor: Ceder, Gerbrand.
- 학위논문주기
- Thesis (Ph.D.)--University of California, Berkeley, 2025.
- 초록/해제
- 요약Artificial intelligence (AI) is increasingly shifting the paradigm of scientific discovery to accelerate research and solve real-world scientific challenges. While ab initio quantum mechanical simulation methods, such as density functional theory (DFT), offer the theoretical foundation to investigate material and chemical science problems at the atomic and electronic levels, their computational demands limit their applicability in both spatial and temporal scales. On the other hand, machine learning interatomic potentials (MLIPs) have enabled the opportunity to scale up quantum mechanical simulations to realistic scales by leveraging machine learning (ML) and neural networks (NNs) to learn and emulate atomic interactions. Recent advancements have led to foundation potentials (FPs) that are pre-trained on comprehensive materials datasets, providing opportunities for universal interatomic potential and foundation ML models.This thesis discusses the datasets, algorithms, benchmarks, and applications of MLIPs and FPs. We introduce the Crystal Hamiltonian Graph neural-Network (CHGNet), a graph neural network (GNN) based FP that learns and predicts the universal potential energy surface (PES) with the incorporation of atomic charge information. CHGNet is pre-trained on the Materials Project Trajectory (MPtrj) dataset, a comprehensive collection of over 1.5 million DFT calculations that spans the whole periodic table. We will demonstrate the application of CHGNet in modeling multiple energy-storage materials, including cathodes such as LixMnO2, LixMn0.8Ti0.1O1.9F0.1, LixFePO4, and solid electrolytes such as Li3La3Te2O12 and Li2ZrCl6. We highlight the insights and design principles that can be learned from the charge-informed large-scale simulations enabled by CHGNet.To further extend the simulation scale of MLIPs, we present DistMLIP, an efficient multi-GPU distributed inference platform for MLIP simulations. By partitioning the atomic graphs across multiple GPUs with DistMLIP, we show that current FPs can perform million-atom-scale simulations with a few GPUs.Benchmarking MLIP's accuracy is critical to ensure its modeling reliability. We discuss the rigorous benchmark study undertaken by Matbench Discovery to test FPs' ability to predict crystal stability through calculating bulk materials' energies. To further assess the FP's performance in out of distribution (OOD) PES regions, we present a series of benchmarks on materials modeling tasks such as surface and defect calculations. We found a consistent PES softening effect in all FPs and modeling tasks, characterized by energy and force under predictions. The PES softening issue behavior originates primarily from the systematically underpredicted PES curvature, which derives from the biased sampling of near-equilibrium atomic arrangements in FP pre-training datasets.We highlight the need for comprehensively sampled datasets to improve the reliability of FPs and avoid failure mechanisms such as the PES softening issue we identified. To address this concern, we will discuss the development of MatPES, a next-generation foundational materials dataset with improved PES sampling and high-fidelity DFT calculations. We will also discuss the open-sourced MLIP library, MatGL, that fosters straightforward adoption of material AIs. We hope these efforts discussed in this thesis present an enabling step towards reliable AI for materials discovery at scale and can continuously support advancements in materials and AI research.
- 일반주제명
- Computational chemistry
- 일반주제명
- Materials science
- 일반주제명
- Engineering
- 키워드
- Li-ion battery
- 키워드
- Machine learning
- 기타저자
- University of California, Berkeley Materials Science & Engineering
- 기본자료저록
- Dissertations Abstracts International. 87-04B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017359113
■00520260202104836
■006m o d
■007cr#unu||||||||
■020 ▼a9798297600669
■035 ▼a(MiAaPQ)AAI32171584
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a542
■1001 ▼aDeng, Bowen.
■24510▼aTowards Reliable AI for Materials Discovery at Scale
■260 ▼a[Sl]▼bUniversity of California, Berkeley▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a151 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-04, Section: B.
■500 ▼aAdvisor: Ceder, Gerbrand.
■5021 ▼aThesis (Ph.D.)--University of California, Berkeley, 2025.
■520 ▼aArtificial intelligence (AI) is increasingly shifting the paradigm of scientific discovery to accelerate research and solve real-world scientific challenges. While ab initio quantum mechanical simulation methods, such as density functional theory (DFT), offer the theoretical foundation to investigate material and chemical science problems at the atomic and electronic levels, their computational demands limit their applicability in both spatial and temporal scales. On the other hand, machine learning interatomic potentials (MLIPs) have enabled the opportunity to scale up quantum mechanical simulations to realistic scales by leveraging machine learning (ML) and neural networks (NNs) to learn and emulate atomic interactions. Recent advancements have led to foundation potentials (FPs) that are pre-trained on comprehensive materials datasets, providing opportunities for universal interatomic potential and foundation ML models.This thesis discusses the datasets, algorithms, benchmarks, and applications of MLIPs and FPs. We introduce the Crystal Hamiltonian Graph neural-Network (CHGNet), a graph neural network (GNN) based FP that learns and predicts the universal potential energy surface (PES) with the incorporation of atomic charge information. CHGNet is pre-trained on the Materials Project Trajectory (MPtrj) dataset, a comprehensive collection of over 1.5 million DFT calculations that spans the whole periodic table. We will demonstrate the application of CHGNet in modeling multiple energy-storage materials, including cathodes such as LixMnO2, LixMn0.8Ti0.1O1.9F0.1, LixFePO4, and solid electrolytes such as Li3La3Te2O12 and Li2ZrCl6. We highlight the insights and design principles that can be learned from the charge-informed large-scale simulations enabled by CHGNet.To further extend the simulation scale of MLIPs, we present DistMLIP, an efficient multi-GPU distributed inference platform for MLIP simulations. By partitioning the atomic graphs across multiple GPUs with DistMLIP, we show that current FPs can perform million-atom-scale simulations with a few GPUs.Benchmarking MLIP's accuracy is critical to ensure its modeling reliability. We discuss the rigorous benchmark study undertaken by Matbench Discovery to test FPs' ability to predict crystal stability through calculating bulk materials' energies. To further assess the FP's performance in out of distribution (OOD) PES regions, we present a series of benchmarks on materials modeling tasks such as surface and defect calculations. We found a consistent PES softening effect in all FPs and modeling tasks, characterized by energy and force under predictions. The PES softening issue behavior originates primarily from the systematically underpredicted PES curvature, which derives from the biased sampling of near-equilibrium atomic arrangements in FP pre-training datasets.We highlight the need for comprehensively sampled datasets to improve the reliability of FPs and avoid failure mechanisms such as the PES softening issue we identified. To address this concern, we will discuss the development of MatPES, a next-generation foundational materials dataset with improved PES sampling and high-fidelity DFT calculations. We will also discuss the open-sourced MLIP library, MatGL, that fosters straightforward adoption of material AIs. We hope these efforts discussed in this thesis present an enabling step towards reliable AI for materials discovery at scale and can continuously support advancements in materials and AI research.
■590 ▼aSchool code: 0028.
■650 4▼aComputational chemistry
■650 4▼aMaterials science
■650 4▼aEngineering
■653 ▼aLi-ion battery
■653 ▼aMachine learning
■653 ▼aMachine learning interatomic potentials
■653 ▼aMaterials modeling
■653 ▼aSolid state electrolyte
■690 ▼a0794
■690 ▼a0800
■690 ▼a0219
■690 ▼a0537
■71020▼aUniversity of California, Berkeley▼bMaterials Science & Engineering.
■7730 ▼tDissertations Abstracts International▼g87-04B.
■790 ▼a0028
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17359113▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


