CDS&E: Collaborative Research: Designing New Zintl Phases with Motif-based Learning and Ab Initio Methods

该奖项支持一种有用但尚未开发的材料类型的计算和数据科学设计，这种材料称为Zintl相，具有适用于热电、电池和光伏的独特性能。随着计算资源的扩展、算法的进步和数据库的增长，具有定制属性的材料的数据驱动设计变得越来越可行。然而，在实践中，探索巨大的搜索空间仍然具有挑战性，数据驱动模型仍然面临精度和可解释性有限的困难。在这个项目中，研究小组将研究一个中心假设，即某些类别的材料（如Zintl相）的材料特性通过结构内原子群之间的相互作用而不是单个原子之间的相互作用更准确地建模。这种物理直觉将用于构建一套新的方法，将先进的计算机模拟与最先进的机器学习（ML）模型集成在一起，以实现快速的逆向材料设计。除了提高准确性和速度外，这些方法还将用于解释和科学地理解材料中原子群之间的重要相互作用，这些相互作用有助于实现所需的特性。研究小组将使用这些方法来发现新的Zintl相，并与实验合作者一起验证他们的发现。计算材料科学家和计算机科学家组成的跨学科研究团队能够很好地推进这些具有挑战性的基础和技术问题的知识。该项目还支持培养研究生和本科生，使他们成为材料与数据科学交叉领域跨学科研究的下一代领导者。在此背景下计划的活动包括开发跨学科研讨会和关于数据支持材料发现的讲习班，以及将材料应用于ML课程的课程材料。该项目还将通过一个关于材料科学机器学习的播客来支持公众参与。技术概述：该奖项支持基于计算和数据科学的Zintl相材料的逆向设计，该材料可应用于热电、电池和光伏等领域。随着计算能力和方法的进步，新材料的计算设计变得越来越可行。然而，探索具有数百万种化合物的大型化学空间对于从头算方法来说仍然是难以计算的。机器学习（ML）已经成为加速此类大型探索的一种手段，但流行的ML方法在准确性和可解释性方面仍然存在局限性。在这个项目中，研究小组将为材料设计开发新的计算方法，其核心思想是通过结构中基元（原子群）之间的相互作用而不是单个原子之间的相互作用来更好地模拟晶体特性。该团队将把这个想法应用于设计新的Zintl相，其中结构图案强烈影响功能特性。Zintl相是一类未被开发的化合物，提供了发现和设计新相的机会。该项目的最终目标是发现具有新成分和结构类型的Zintl相，并了解它们的成分，结构基元和功能特性之间的关系，从而可以对具有定制属性的Zintl相进行反向设计。该团队将使用化学装饰和类似乐高的组装来“构建”Zintl相，并开发分层图表示算法，以自动学习结构图案和功能属性之间的关系。该奖项反映了美国国家科学基金会的法定使命，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。

NONTECHNICAL SUMMARYThis award supports computational and data-science-enabled design of a useful but under-explored class of materials called Zintl phases, which have unique properties suitable for thermoelectrics, batteries, and photovoltaics. Data-driven design of materials with tailored properties has become increasingly viable with the expansion of computing resources, algorithmic advances, and growing databases. However, it remains challenging in practice to explore vast search spaces, and data-driven models still face difficulties with limited accuracy and interpretability. In this project, the research team will study the central hypothesis that materials properties for certain classes of materials such as Zintl phases are more accurately modeled through interactions between groups of atoms within a structure rather than interactions between individual atoms. This physical intuition will be used to build a new suite of methods integrating advanced computer simulations with state-of-the-art machine learning (ML) models for enabling fast inverse materials design. In addition to improving accuracy and speed, these methods will be designed to enable interpretation and scientific understanding of the important interactions between groups of atoms in the material that contribute to desired properties. The research team will use these methods to discover new Zintl phases and work with experimental collaborators to validate their discoveries. The interdisciplinary research team of computational materials scientists and computer scientists is well positioned to advance knowledge in these challenging problems of fundamental and technological interest. This project also supports training of graduate and undergraduate students to be the next generation of leaders in interdisciplinary research at the intersection of materials and data science. Activities planned within this context include the development of cross-disciplinary seminars and workshops on data-enabled materials discovery and course materials introducing materials applications into ML courses. The project will also support public engagement through a podcast on ML for materials science.TECHNICAL SUMMARYThis award supports computational and data science-enabled inverse design of a class of materials called Zintl phases that have applications in thermoelectrics, batteries, and photovoltaics, among others. Computational design of new materials has become increasingly viable with advances in computing power and methodologies. However, exploration of large chemical spaces with millions of compounds is still computationally intractable for ab-initio methods. Machine learning (ML) has emerged as a means to accelerate such large explorations, but popular ML methods still have limitations in accuracy and interpretability. In this project, the research team will develop novel computational methods for materials design focused on the central idea that crystal properties are better modeled through interactions between motifs (groups of atoms) within a structure rather than between individual atoms. The team will apply this idea to design new Zintl phases, where the structural motifs strongly influence functional properties. Zintl phases are an under-explored class of compounds, presenting opportunities to discover and design new phases. The ultimate goal of this project is to discover Zintl phases with new compositions and structure types and learn relationships between their compositions, structural motifs, and functional properties that will allow inverse design of Zintl phases with tailored properties. The team will use chemical decorations and Lego-like assembly to "construct" Zintl phases and develop hierarchical graph representation algorithms to automatically learn the relationships between structural motifs and functional properties.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.