CAREER: Decoding the Structure and Energy Landscape of Isostatic Glasses by Machine Learning and Enhanced Sampling

该职业奖支持玻璃科学领域的研究和教育活动。当冷却到熔点以下时，液体通常会变成晶体。作为一种替代途径，如果淬火足够快，液体可以绕过结晶而进入过冷液态。在玻璃化转变温度下，液体变得非常粘稠，其流动时间最终超过了观测时间。在这一点上，因为他们不能流动，他们冻结成固体玻璃。如果淬火得足够快，几乎整个元素周期表都能形成玻璃，但液体在淬火时结晶或形成玻璃的倾向，即形成玻璃的能力，取决于它的成分。破译液体的原子组成和结构如何控制它们的玻璃形成能力一直是玻璃科学的基础，迄今为止在很大程度上仍未解决。为此，玻璃的原子结构可以基于网络拓扑的概念来描述，网络拓扑是数学的一个分支，研究节点的连通性，在这种情况下是原子，如何控制网络的性质，即玻璃原子结构。有趣的是，具有最佳网络拓扑结构的玻璃，称为等静力，往往具有独特的性能，包括最大的玻璃形成能力。然而，网络拓扑结构和玻璃形成能力之间联系的本质仍然是未知的。为了解决这些问题，PI试图解码网络拓扑结构和玻璃形成能力之间的关系，因此，询问玻璃的本质，并努力理解等静力玻璃特征的异常性质的起源。这将通过分子动力学模拟、增强采样和机器学习活动的结合来完成，这些活动紧密结合在一起，相互通知和推进。该项目将专注于硫化物玻璃，这是3D Xpoint相变随机存取存储器的基础材料。通过每个单元存储3位而不是单个位，这种技术被广泛认为将构成非易失性存储器的未来。这些研究任务得到若干教育活动的补充和支持。PI将开发一系列跨学科课程，为学生提供横向知识，这是开发综合解决方案以应对全球复杂现实挑战的关键技能。PI还将利用该项目促进基于数据的适应性学习，这为摆脱不反映复杂认知过程的“一刀切”学习方式提供了独特的机会。为了解决工科的性别不平衡问题，PI将开发针对女性学生的本科研究项目，以提高她们在研究生院的留校率。该职业奖支持研究和教育活动，以促进对结构玻璃的基本理解。基于它们的网络拓扑结构，当拓扑原子间约束的数量分别较低、较高或等于原子自由度的数量时，结构玻璃可以被分类为柔性（欠约束）、应力刚性（过度约束）和等静力（刚性，但无应力）。有趣的是，最佳约束等静力玻璃往往表现出异常性质，包括最佳玻璃形成能力和低松弛倾向。这种等静力玻璃被认为存在于中间相中，其中玻璃自组织表现出几乎可逆的玻璃化转变。然而，中间阶段的存在和起源仍然存在争议，到目前为止还没有发现这一阶段的结构特征。此外，均衡玻璃的异常性质的起源仍然是难以捉摸的。为了解决这些知识上的空白，PI试图解码网络拓扑结构、松弛倾向和玻璃形成能力之间的关系。为此，PI将采用综合方法，其中经典分子动力学，密度泛函理论，增强采样和机器学习活动相互通知和推进。首先，将训练机器学习力场来模拟典型的Ge-As-Se硫系玻璃。其次，该力场将用于解码拓扑与松弛倾向之间的关系，并询问中间相的存在性。第三，将使用机器学习寻求自组织的结构特征。最后，增强采样技术将用于表征具有不同网络拓扑结构的玻璃的能量景观。通过物理和数据驱动建模技术的协同结合，本研究将揭示网络拓扑结构、松弛倾向和玻璃形成能力之间的关系如何在能量景观的地形中编码。本研究还将探讨硫系玻璃中中间相和结构自组织的存在、起源和特征。该项目还支持一项广泛的教育计划。首先，PI将引入一个连接凝聚态物质、土木工程和计算机科学的跨学科项目。其次，本文开发的机器学习技术将用于在课堂上实现数据知情的主动学习。第三，将利用与工业伙伴的积极合作，使学生接触工业环境和企业文化。第四，PI将制定针对女性的本科研究计划，以提高她们在研究生院的保留率。最后，PI将与当地高中教师合作开发演示模块，其中包括虚拟现实，结构桁架套件和糖玻璃结晶等活动，以支持K-12学生的材料科学推广。通过这些活动，该项目将有助于支持美国的玻璃科学。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。

NON-TECHNICAL SUMMARYThis CAREER award supports research and education activities in the field of glass science. When cooled below their melting point, liquids usually turn into crystals. As an alternative route, liquids can bypass crystallization and enter a supercooled liquid state if quenched fast enough. At the glass transition temperature, liquids become so viscous that their flowing time eventually exceeds the observation time. At this point, since they are unable to flow, they freeze into solid glasses. Although virtually the entire periodic table can form a glass if quenched fast enough, the propensity for a liquid to crystallize or form a glass upon quenching, its glass-forming ability, depends on its composition. Deciphering how the atomic composition and structure of liquids govern their glass-forming ability has been at the very foundation of glass science and has remained largely unresolved thus far. To this end, the atomic structure of glasses can be described based on the concept of network topology, that is, the branch of mathematics that studies how the connectivity of nodes, in this case - the atoms, govern the properties of networks, the glass atomic structure. Interestingly, glasses exhibiting an optimal network topology, called isostatic, tend to feature unique properties, including maximum glass-forming ability. However, the nature of the linkages between network topology and glass-forming ability remains largely unknown.To address these questions, the PI seeks to decode the relationship between network topology and glass-forming ability, and so, interrogates the very nature of a glass and works toward understanding the origin of the anomalous properties featured by isostatic glasses. This will be accomplished by a combination of molecular dynamics simulations, enhanced sampling, and machine learning activities, which are closely integrated to inform and advance each other. This project will focus on chalcogenide glasses, which are the base material for 3D Xpoint phase-change random access memory. By storing 3 bits per cell instead of a single bit, this technology is widely believed to constitute the future of non-volatile memory.These research tasks are complemented and enabled by several educational activities. The PI will develop a series of cross-disciplinary courses to equip students with transversal knowledge, which is a key skillset to develop integrated solutions that globally address complex real-world challenges. The PI will also leverage this project to promote data-based adaptive learning, which offers a unique opportunity to move away from the "one size fits all" way of learning that does not reflect complex cognitive processes. To address the gender imbalance in engineering, the PI will develop an undergraduate research program targeted to female students, which aims to enhance their retention in graduate school.TECHNICAL SUMMARY This CAREER award supports research and education activities to advance fundamental understanding of structural glasses. Based on their network topology, structural glasses can be classified as flexible (underconstrained), stressed-rigid (overconstrained), and isostatic (rigid, but free of stress) when the number of topological interatomic constraints is lower, higher, or equal to the number of atomic degrees of freedom, respectively. Interestingly, optimally-constrained isostatic glasses tend to exhibit anomalous properties, including optimal glass-forming ability and low propensity for relaxation. Such isostatic glasses have been suggested to exist within an intermediate phase, wherein the glass self-organizes to exhibit a nearly-reversible glass transition. However, the existence and origin of the intermediate phase remain debated and no structural signature of this phase has been revealed thus far. Further, the origin of the anomalous properties of isostatic glasses remains elusive. To address these gaps in knowledge, the PI seeks to decode the relationship between network topology, propensity for relaxation, and glass-forming ability.To this end, the PI will adopt an integrated approach, wherein classical molecular dynamics, density functional theory, enhanced sampling, and machine learning activities mutually inform and advance each other. First, a machine-learned forcefield will be trained to simulate archetypical Ge-As-Se chalcogenide glasses. Second, this forcefield will be used to decode the relationship between topology and propensity for relaxation, and interrogate the existence of the intermediate phase. Third, a structural signature of self-organization will be sought using machine learning. Last, enhanced sampling techniques will be used to characterize the energy landscape of glasses with varying network topologies. By synergistically combining physics- and data-driven modeling techniques, this research will reveal how the relationship between network topology, propensity for relaxation, and glass-forming ability is encoded in the topography of the energy landscape. This study will also interrogate the existence, origin, and signatures of the intermediate phase and structural self-organization in chalcogenide glasses. This project also supports an extensive education plan. First, the PI will introduce a cross-disciplinary program connecting condensed matter, civil engineering, and computer science. Second, the machine learning techniques developed herein will be used to enable data-informed active learning in classrooms. Third, active collaborations with industrial partners will be leveraged to expose students to industrial environments and corporate cultures. Fourth, the PI will develop an undergraduate research program targeted to females, so as to enhance their retention in graduate school. Finally, the PI will collaborate with local high school teachers to develop demonstration modules, with activities involving virtual reality, structural truss kit set, and sugar glass crystallization to support material science outreach to K-12 students. Through these activities, the project will contribute to supporting glass science in the United States.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.