Atomically dispersed amorphous catalysts: ab initio computational tools for a new frontier

该项目旨在开发和应用计算方法来了解非晶支撑材料上分离金属原子位点的催化性能。计算结果将与菲利普斯石油公司乙烯聚合催化剂的实验数据进行比较，该催化剂60年来一直是工业催化剂的中坚力量，尽管对其活性位点和机理存在长期问题。该结果不仅将为菲利普斯催化剂的潜在改进提供具体信息，而且将改进理论工具，以理解活性由非晶载体上一小部分高活性金属位点主导的广泛类型的催化剂。相关的教育和推广项目将提供给各个层次的学生，包括一个让高中生参与科学追求的游戏。该研究将开发计算技术，以确定在非晶支撑材料上的一组孤立金属位点中，使某些催化位点相对于其他位点具有高活性的特定性质。这项工作特别侧重于支持无定形二氧化硅（菲利普斯催化剂）的铬，这是一个广泛的表征数据体。计算方法将结合机器学习技术和罕见事件方法来分析站点的分布，以预测结构和活动之间的关系。具体来说，通过从头计算训练的机器学习方法将学习活性如何与二氧化硅表面分离的铬种的局部结构环境相关。非玻尔兹曼采样技术将确保具有异常高活度（低活化能）的稀有但重要的位点得到充分采样，以确保准确的位点平均动力学性质。该方法提供了一种新的“重要性学习”策略，可广泛用于建立活性位点分布模型，识别高活性位点的关键特征，以及在非晶载体上设计更好的原子分散催化剂。更广泛的教育和推广贡献包括开发和公开分享支持重要性学习方法的“插件”，机器学习工具的虚拟现实可视化，以及面向高中生的相关游戏，他们将有机会与机器学习算法竞争，设计高活性催化位点。

The project aims to develop and apply computational methods to understand the catalytic properties of isolated metal atom sites on amorphous support materials. The computational results will be compared to experimental data on the Phillips Petroleum ethylene polymerization catalyst which has been a workhorse industrial catalyst for 60 years despite longstanding questions about the active sites and the mechanism. The results will not only provide information specific to potential improvements in the Phillips catalyst, but will improve theoretical tools for understanding a broad class of catalysts where the activity is dominated by a small fraction of highly active metal sites on the amorphous support. Related educational and outreach programs will be offered to students at all levels, including a game to engage high school students in scientific pursuits.The study will develop computational techniques to identify the specific properties that make certain catalytic sites highly active relative to others among an ensemble of isolated metal sites on an amorphous support material. The work specifically focuses on chromium supported on amorphous silica (the Phillips catalyst) for which a broad body of characterization data is available. The computational approach will combine machine learning techniques and rare events methods for analyzing the distribution of sites to predict relationships between structure and activity. Specifically, machine learning methods trained by ab initio calculations will learn how activity is related to the local structural environments of the isolated chromium species on the silica surface. Non-Boltzmann sampling techniques will ensure that rare but important sites with unusually high activities (low activation energies) are adequately sampled to ensure accurate site-averaged kinetic properties. The combined approach provides a new "importance learning" strategy that can be broadly used to build models of active site distributions, identify critical characteristics of highly active sites, and engineer better atomically dispersed catalysts on amorphous supports. Broader educational and outreach contributions include the development and public sharing of "plug-ins" that support the importance learning approach, virtual reality visualization of the machine learning tools, and a related game for high school students in which they will have an opportunity to compete with the machine learning algorithm to design highly active catalytic sites.