CDS&E: Inferring Lattice Dynamics from Temporal X-ray Diffraction Data

新兴的x射线散射实验技术提供了探测原子结构和材料物理性质之间相关性的能力，具有原子尺度（约十亿分之一米）的灵敏度和万亿分之一秒的时间分辨率。这些散射实验所获得的图像包含了环、斑、晕等视觉特征，这些特征编码了原子结构及其时间演化的详细信息。然而，相应的数据集非常大，因此，人工分析具有许多不确定性的数据集不能完全由人类专家来完成。该奖项支持研究和教育活动，以开发人工智能技术，从x射线散射实验中挖掘数据，以检测材料在高压、高温或应变等极端条件下的相变和塑性变形的原子尺度机制。实现对控制原子排列和运动的机制的基本理解对于确定在极端条件下形成具有所需性质和行为的新物质的新途径至关重要。该项目还将为本科生和研究生提供计算材料科学、高级原子水平结构/化学表征、分子动力学模拟和人工智能技术方面的多学科培训。该项目将为人工智能应用于材料科学的新课程材料和模块的设计提供信息。pi还将参与到K-12学生的外展活动中，旨在扩大未被充分代表的群体在科学、技术、工程和数学方面的参与，并提高对纳米技术和材料科学的认识。该奖项支持旨在开发自动化深度学习计算机视觉技术的研究和教育活动，以挖掘x射线衍射（XRD）数据，以识别晶体结构，并检测极端条件下相变和塑性变形的晶格级机制。在非常高的压力、温度或应变率下，当晶格变化或新相的出现事先未知时，分析数十亿次XRD测量的大量快照数据集变得不准确，甚至完全失败。为了克服这一挑战，研究团队将利用一系列新颖而先进的技术，包括多模态融合、重建、时空建模、弱监督、领域自适应和可视化，以实现以下目标：1)生成静态和时间合成的一维XRD图和二维XRD图；2)开发用于晶体结构静态和时间分类的深度学习模型；3)开发用于解释和证明深度学习模型和预测的解释技术；4)对大型实验数据的领域自适应。这种深度学习技术的成功发展将导致在没有先验知识的极端条件下对材料中未知现象的更深入理解。该项目还将为本科生和研究生提供计算材料科学、高级原子级结构/化学表征、分子动力学模拟和深度学习技术方面的多学科培训。该项目将为材料科学应用深度学习的新课程材料和模块的设计提供信息。pi还将参与到K-12学生的外展活动中，旨在扩大未被充分代表的群体在科学、技术、工程和数学方面的参与，并提高对纳米技术和材料科学的认识。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 SUMMARYEmerging X-ray scattering experimental techniques provide the capability to probe correlations between atomic structure and physical properties of materials with atomic-scale (around one-billionth of a meter) sensitivity and with one trillionth of a second time resolution. The images obtained in these scattering experiments contain visual features, such as rings, spots, and halos, which encode detailed information about the atomic structure and its time evolution. However, the corresponding data sets are enormously large, and therefore, manually analyzing them with many uncertainties cannot be solely performed by human experts. This award supports research and educational activities to develop artificial intelligence techniques to mine data from X-ray scattering experiments to detect atomic-scale mechanisms of phase transformations and plastic deformation in materials when they are subjected to extreme conditions like high pressure, temperature or strain. Achieving a fundamental understanding of the mechanisms that govern the arrangement and motion of atoms is crucial for identifying new pathways of forming new matter with desired properties and behavior at extreme conditions. This project will also provide multidisciplinary training for undergraduate and graduate students in computational materials science, advanced atomic-level structural/chemical characterization, molecular dynamics simulations, and artificial intelligence techniques. The project will inform the design of new course material and modules on artificial intelligence applied to materials science. The PIs will also be involved in outreach to K-12 students aimed at broadening participation of underrepresented groups in science, technology, engineering, and mathematics and raise awareness of nanotechnology and materials science.TECHNICAL SUMMARYThis award supports research and educational activities aimed at developing automated deep-learning computer vision techniques to mine x-ray diffraction (XRD) data to identify crystal structures and detect lattice-level mechanisms responsible for phase transformation and plastic deformation under extreme conditions. At very high pressures, temperatures, or strain rates when lattice variations or occurrence of new phases are not known a priori, analyzing vast datasets of snapshots from billions of XRD measurements become inaccurate, or fail completely. To overcome this challenge, the research team will leverage a series of novel and advanced techniques, including multimodal fusion, reconstruction, space-time modeling, weak supervision, domain adaptation, and visualization to achieve the following objectives: 1) Generation of static and temporal synthetic one-dimensional XRD patterns and two-dimensional XRD images, 2) Development of deep learning models for static and temporal classification of crystal structures, 3) Development of interpretation techniques for explanation and justification of deep learning models and predictions, and 4) Domain adaptation to large experimental data. The successful development of such deep learning techniques will lead to deeper understanding of unknown phenomena in materials under extreme conditions when no prior knowledge is available.This project will also provide multidisciplinary training for undergraduate and graduate students in computational materials science, advanced atomic-level structural/chemical characterization, molecular dynamics simulations, and deep learning techniques. The project will inform the design of new course material and modules on applied deep learning for materials science. The PIs will also be involved in outreach to K-12 students aimed at broadening participation of underrepresented groups in science, technology, engineering, and mathematics and raise awareness of nanotechnology and materials science.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.