EAGER: SSMCDAT2023: Revealing Local Symmetry Breaking in Intermetallics: Combining Statistical Mechanics and Machine Learning in PDF Analysis

第一部分：非技术性总结这个奖项是在一个热切的提议上颁发的。它支持在利哈伊大学举行的SSMCDAT 2023数据马拉松上推进的一个项目的进展。这个急切的项目进一步加深了对半导体结构的了解，特别是那些原子无序程度很高的半导体。半导体对现代电子来说至关重要，它们的物理性质，如导电性和带隙，在很大程度上取决于它们的原子排列。尽管在过去的一个世纪里在研究有序晶体方面取得了重大进展，但理解无序晶体的结构仍然是一个挑战。理解这一点的关键是成分和合成方法等因素如何影响原子结构，从而影响电子和热传输等性质。这个项目使用尖端的人工智能技术，特别是对称感知神经网络，来预测无序材料中原子之间的作用力。这些预测有助于模拟实验测量的结果，提供对这些复杂结构的洞察。考虑到这些无序材料的内在复杂性，该项目还涉及开发在原子水平上可视化这些结构的创新方法，使研究人员能够识别半导体结构和电子性质之间的模式。第2部分：技术总结金属间化合物和合金中成分、加工条件和材料性质之间的预测关系的建立受到理解局部结构的困难的阻碍。在金属间化合物材料中，局域结构既包括晶格上的化学瓦片(即基元)，也包括由不对称配位环境引起的原子扭曲。即使这样的结构具有高保真度，局部结构的多样性仍然具有可视化的挑战性。因此，衡量和了解当地结构仍然是关键挑战，使目前的金属间材料开发工作主要是经验性的。这个急切的项目通过一种综合的计算-实验方法来解决这些挑战。Ge(1-x)MnxTe体系具有显著的溶解性、强的中子散射和随温度和成分的连续相变，可作为模型体系。从不同成分的金属间化合物中收集的中子对分布函数测量(PDF)获得了实验上的见解。对这些PDF测量进行拟合的计算洞察力取决于(I)用于分子动力学模拟的基于等变神经网络的力场的最新进展，以及(Ii)构型和振动能量的稳健统计力学处理。这种拟合过程仍然适用于金属间化合物或合金中的潜在结合能，这意味着可以仔细检查所产生的结构，以了解它们的局部结构分布。这个项目的一个关键部分是对由此产生的结构扭曲和粘合进行特色化。特色化优化、无监督学习和可视化发展将允许深入了解加工条件和成分对这些材料的本地结构的影响。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

PART 1: NON-TECHNICAL SUMMARYThis award is made on an EAGER proposal. It supports progress on a project advanced at the SSMCDAT 2023 Datathon held at Lehigh University. This EAGER project furthers the understanding of the structure of semiconductors, especially those with a high degree of atomic disorder. Semiconductors are critical to modern electronics, with their physical properties, like conductivity and band gap, largely depending on their atomic arrangements. Despite significant progress in studying well-ordered crystals over the past century, understanding the structure of disordered crystals remains a challenge. Key to this understanding is how factors such as composition and synthesis methods affect atomic structure and, consequently, properties like electronic and thermal transport. This project employs cutting-edge artificial intelligence techniques, specifically symmetry-aware neural networks, to predict the forces between atoms in disordered materials. These predictions aid in modeling the results of experimental measurements, offering insight into these complex structures. Given the inherent complexity of these disordered materials, the project also involves developing innovative ways to visualize these structures at an atomic level, enabling researchers to identify patterns between semiconductor structure and electronic properties.PART 2: TECHNICAL SUMMARYEstablishing predictive relationships between composition, processing conditions, and material properties in intermetallics and alloys has been hampered by difficulties in understanding local structure. In an intermetallic material, the local structure includes both the chemical tiling on the lattice (i.e., motifs) and atomic distortions arising from asymmetric coordination environments. Even with high fidelity determination of such structures, the diversity of local structures remains challenging to visualize. As such, both measurement and understanding of local structure remain key challenges, making current intermetallic materials development efforts primarily empirical. This EAGER project addresses these challenges with an integrated computational-experimental approach. The Ge(1-x)MnxTe system serves as a model system due to its significant solubility, strong neutron scattering, and continuous phase transition with temperature and composition. Experimental insights are gleaned from neutron pair distribution function measurements (PDF) collected from intermetallics with various compositions. Computational insights into fitting these PDF measurements hinge on (i) recent advancements in equivariant neural net-based force fields for molecular dynamics simulations and (ii) a robust statistical mechanics treatment of configurational and vibrational energies. This fitting procedure remains true to the underlying bonding energetics within the intermetallic or alloy, meaning the resulting structures can be scrutinized for their distribution of local structures. A key part of this project is the featurization of the resulting structural distortions and bonding. Featurization optimization, unsupervised learning, and visualization developments will allow insights into the impact of processing conditions and composition on the local structure of these materials.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.