CDS&E: Machine learning enabled modelling of dynamic nanoparticle catalysts

CDS

• 批准号: 2152767
• 负责人: Philippe Sautet
• 金额: $ 36.43万
• 依托单位: University of California-Los Angeles
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2022
• 资助国家: 美国
• 起止时间: 2022-09-01 至 2025-08-31
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2152767&HistoricalAwards=false
• 关键词: CDS& Machine; learning; enabled; modelling

Catalysis enables chemical and fuels manufacturing by promoting efficient energy and resource utilization. As widely practiced in industry, catalysis is an inherently complex process, which has led to heavy reliance on trial-and-error methods for catalyst discovery and optimization. Advances in computational and data science methods are seeing increasing application in catalyst discovery and design. The project extends those methods to a range of catalyst particle sizes that are common in both industrial and environmental applications, yet challenging to simulate because of 1) the large number of atoms which must be modeled, 2) the “massive” combinatorial space of their structural configurations, and 3) dynamical changes in structure that occur in response to changes in reaction conditions. In contrast to previous modeling approaches that treat small particles as static, perfect polyhedrons, the project develops methods that consider the effects of dynamic rearrangements and metastable structures on dehydrogenation and hydrogenolysis reactions as catalyzed by platinum and nickel nanoparticles containing 20 to 200 atoms. To that end, the project employs several novel data science methods to model the vast computational space needed to predict relationships between dynamic catalyst structures and chemical reactivity. Beyond the technical aspect, the project offers educational and outreach activities focused on underrepresented students at the undergraduate and high-school levels.The project is built on the hypothesis that a transition metal nanoparticle, for example based on Pt or Ni, presents several low energy metastable isomers, representing accessible “defect-type” structures, that might fully dominate the catalytic activity of the particle. The main difficulty for atomistic modelling of metallic nanoparticles is the massive combinatorial space of their structural configurations. Those challenges will be addressed by exploiting data science methods leading to generalized models linking the structure and reactivity of nanoparticles in dynamical situations. Data science methods will be used at two places in the project, 1) interatomic potential fitting, and 2) identification of structural motifs on nanoparticle surfaces. The first step involves generation of an accurate inter-atomic potential using Neural Networks based on the investigator’s previous experience modeling small clusters. The various configurations of the surface atoms on the nanoparticle will be explored using basin-hoping algorithms, in a grand canonical approach to handle a variable number of adsorbates. This will allow us to efficiently explore the extremely diverse structures for Pt and Ni particles of ~20-200 atoms, with realistic adsorbate coverage. The obtained large structure database will be used to extract local structure descriptors and learn the statistical distribution of local structural motifs, using pattern recognition algorithms. This distribution of local motifs is key to analyze the catalytic activity. The educational and outreach aspects involve partnering with the UCLA Center for Excellence in Engineering and Diversity (CEED) to involve both high-school and undergraduate URM students in the research. A second initiative features a one-day workshop for high-school teachers on the UCLA campus, targeting Latino schools in central Los Angeles. The program will illustrate how computational chemistry can provide key insights on how catalysts work at the atomic scale.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.

催化通过促进有效的能源和资源利用使化学和燃料制造成为可能。 正如在工业中广泛实践的那样，催化是一个固有的复杂过程，这导致严重依赖试错法来发现和优化催化剂。 计算和数据科学方法的进步在催化剂发现和设计中的应用越来越多。 该项目将这些方法扩展到工业和环境应用中常见的一系列催化剂颗粒尺寸，但模拟具有挑战性，因为1）必须建模的大量原子，2）其结构配置的“大规模”组合空间，以及3）响应反应条件变化而发生的结构动态变化。 与以前将小颗粒视为静态完美多面体的建模方法相反，该项目开发了考虑动态重排和亚稳结构对脱氢和氢解反应的影响的方法，这些反应由含有20至200个原子的铂和镍纳米颗粒催化。 为此，该项目采用了几种新的数据科学方法来模拟预测动态催化剂结构和化学反应性之间关系所需的巨大计算空间。 除了技术方面，该项目还提供教育和推广活动，重点关注本科和高中阶段代表性不足的学生。该项目建立在这样一种假设之上：过渡金属纳米颗粒，例如基于Pt或Ni的纳米颗粒，呈现出几种低能亚稳异构体，代表着可接近的“缺陷型”结构，可能完全主导颗粒的催化活性。金属纳米粒子的原子模型的主要困难是其结构配置的大规模组合空间。这些挑战将通过利用数据科学方法来解决，这些方法将导致在动态情况下将纳米颗粒的结构和反应性联系起来的通用模型。 数据科学方法将用于该项目的两个地方，1）原子间势拟合，2）识别纳米颗粒表面的结构基序。 第一步涉及使用神经网络生成精确的原子间势，该神经网络基于研究者先前对小簇建模的经验。纳米颗粒表面原子的各种配置将使用希望盆地算法进行探索，以巨正则方法处理可变数量的吸附物。这将使我们能够有效地探索约20-200个原子的Pt和Ni颗粒的极其多样化的结构，并具有真实的吸附物覆盖范围。所获得的大型结构数据库将用于提取局部结构描述符，并使用模式识别算法学习局部结构基序的统计分布。局部基序的这种分布是分析催化活性的关键。 教育和推广方面涉及与加州大学洛杉矶分校卓越工程和多样性中心（CEED）合作，让高中和本科URM学生参与研究。 第二项举措是在加州大学洛杉矶分校校园为高中教师举办为期一天的讲习班，目标是洛杉矶中部的拉丁美洲学校。该计划将说明计算化学如何提供关于催化剂在原子尺度上如何工作的关键见解。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。