Collaborative Research: Integrating Physics and Generative Machine Learning Models for Inverse Materials Design

合作研究：将物理与生成机器学习模型相结合进行逆向材料设计

• 批准号: 1940118
• 负责人: Jianzhong Wu
• 金额: $ 42.44万
• 依托单位: University of California-Riverside
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2019
• 资助国家: 美国
• 起止时间: 2019-10-01 至 2023-09-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1940118&HistoricalAwards=false
• 关键词: Collaborative; Research; Integrating; Physics; Generative

This project is aimed to address a grand challenge in data-intensive materials science and engineering to find better materials with desired properties, often with the goal to enhance performance in specific applications. This project addresses this grand challenge with a specific focus on finding metal organic framework (MOF) materials that are used to separate gas mixtures and finding better battery materials for energy storage. The PIs will combine theoretical methods from statistical mechanics and condensed-matter physics, and physics-based models, to generate information-rich materials data which is integrated with generative machine learning (ML) algorithms to search a complex chemical design space efficiently and to train deep learning models for fast screening of materials properties. This project will be carried out by a multidisciplinary collaboration involving researchers from physics, materials science and engineering, computer science, and mathematics. The resulting multidisciplinary environment fosters training the next generation data savvy scientists who will engage in collaborative multidisciplinary research. Existing approaches for computational design of metal organic frameworks (MOF) and solid-state electrolyte materials are largely based on screening of known materials or enumerative search of hypothetical materials. This project develops a new approach that integrates first principles calculations, experimental data and abundant data generated by physics-based models to train generalized antagonistic network (GAN) models for efficient search of the materials design space, and to train deep convolutional neural network (DCNN) models for fast and accurate screening of properties of the GAN-generated candidate materials. Additionally, graph-based GAN models will be used for MOF topology exploration and can be applied to other nanomaterials designs. More specifically, the investigators will: 1) develop and exploit physics-based models for fast calculation of properties such as diffusivity, ion conductivity, and mechanical stability; 2) develop generative adversarial network (GAN) models with built-in physics rules for efficient exploration of the chemical design space for both MOF materials and solid electrolytes; 3) use persistence homology and Bravais lattice sequence representations of MOF materials and solid electrolytes, respectively, to build Deep Convolutional Neural Network (DCNN) models for fast and accurate prediction of the physical properties of generated materials; 4) apply high-level quantum-mechanical calculations for verification of discovered materials. Accomplishments from this project will lead to accelerated discovery of novel nanostructured materials for gas separation and energy storage, materials for lithium-ion batteries, novel data-driven scheme for materials design, and theoretical methods enabling implementation of advanced data science techniques. The highly interdisciplinary collaboration will offer students unique opportunities to interact with a variety of disciplines, and training the next-generation scientists with the mindset for multidiscipline collaborations. Educational and outreach activities will be developed and undertaken in conjunction with the proposed research activities.This project is part of the National Science Foundation's Harnessing the Data Revolution (HDR) Big Idea activity, and is jointly supported by HDR and the Division of Materials Research within the NSF Directorate of Mathematical and Physical Sciences.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

该项目旨在解决数据密集型材料科学和工程领域的巨大挑战，寻找具有所需性能的更好材料，通常目标是提高特定应用中的性能。该项目应对这一巨大挑战，重点关注寻找用于分离气体混合物的金属有机框架（MOF）材料以及寻找更好的储能电池材料。 PI 将结合统计力学和凝聚态物理学的理论方法以及基于物理的模型，生成信息丰富的材料数据，这些数据与生成机器学习 (ML) 算法集成，以有效搜索复杂的化学设计空间并训练用于快速筛选材料特性的深度学习模型。该项目将由多学科合作开展，涉及物理学、材料科学与工程、计算机科学和数学的研究人员。由此产生的多学科环境有利于培养下一代精通数据的科学家，他们将参与多学科协作研究。 金属有机框架（MOF）和固态电解质材料的现有计算设计方法主要基于已知材料的筛选或假设材料的枚举搜索。该项目开发了一种新方法，集成了第一原理计算、实验数据和基于物理的模型生成的大量数据，以训练广义对抗网络（GAN）模型以有效搜索材料设计空间，并训练深度卷积神经网络（DCNN） ）用于快速准确筛选 GAN 生成的候选材料特性的模型。此外，基于图的 GAN 模型将用于 MOF 拓扑探索，并可应用于其他纳米材料设计。更具体地说，研究人员将：1）开发和开发基于物理的模型，用于快速计算扩散率、离子电导率和机械稳定性等特性； 2）开发具有内置物理规则的生成对抗网络（GAN）模型，以有效探索MOF材料和固体电解质的化学设计空间； 3）分别使用MOF材料和固体电解质的持久同源性和布拉维晶格序列表示来构建深度卷积神经网络（DCNN）模型，以快速准确地预测所生成材料的物理性质； 4）应用高级量子力学计算来验证已发现的材料。该项目的成就将加速发现用于气体分离和储能的新型纳米结构材料、锂离子电池材料、用于材料设计的新型数据驱动方案以及能够实施先进数据科学技术的理论方法。高度跨学科的合作将为学生提供与各种学科互动的独特机会，并培养具有多学科合作思维的下一代科学家。教育和推广活动将与拟议的研究活动一起制定和开展。该项目是国家科学基金会利用数据革命 (HDR) 大创意活动的一部分，并得到 HDR 和材料研究部的共同支持NSF 数学和物理科学理事会。该奖项反映了 NSF 的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

项目成果

Transferability evaluation of the deep potential model for simulating water-graphene confined system

模拟水-石墨烯约束系统的深电位模型的可迁移性评估

• DOI: 10.1063/5.0153196
• 发表时间: 2023-07
• 期刊: The Journal of Chemical Physics
• 作者: Liu, Dongfei;Wu, Jianzhong;Lu, Diannan
• 通讯作者: Lu, Diannan

Physics-informed Gaussian process regression of in operando capacitance for carbon supercapacitors

碳超级电容器操作电容的基于物理的高斯过程回归

• DOI: 10.1039/d3ya00071k
• 发表时间: 2023-06
• 期刊: Energy Advances
• 作者: Pan, Runtong;Gu, Mengyang;Wu, Jianzhong
• 通讯作者: Wu, Jianzhong

Predicting the Materials Properties Using a 3D Graph Neural Network With Invariant Representation

使用具有不变表示的 3D 图神经网络预测材料属性

• DOI: 10.1109/access.2022.3181750
• 发表时间: 2022-01
• 期刊: IEEE Access
• 影响因子: 3.9
• 作者: Zhang, Boyu;Zhou, Mushen;Wu, Jianzhong;Gao, Fuchang
• 通讯作者: Gao, Fuchang

Massively Parallel GPU-Accelerated String Method for Fast and Accurate Prediction of Molecular Diffusivity in Nanoporous Materials

• DOI: 10.1021/acsanm.1c00727
• 发表时间: 2021-05-05
• 作者: Musen Zhou;Jianzhong Wu
• 通讯作者: Jianzhong Wu

A GPU implementation of classical density functional theory for rapid prediction of gas adsorption in nanoporous materials

经典密度泛函理论的 GPU 实现，用于快速预测纳米多孔材料中的气体吸附

• DOI: 10.1063/5.0020797
• 发表时间: 2020-08
• 期刊: The Journal of Chemical Physics
• 作者: Zhou, Musen;Wu, Jianzhong
• 通讯作者: Wu, Jianzhong