Collaborative Research: EAGER-DynamicData: Probabilistic Analysis of Dynamic X-ray Diffraction Data: Toward Validated Computational Models for Polycrystalline Plasticity

过去二十年见证了针对广泛物理过程的准确和高效计算方法的发展，并将这些模型转变为美国经济所有部门中用于产品设计和开发的常用工具。这一趋势的一个重要例外是在材料科学领域，由于缺乏经过验证的计算模型，在创造新的材料类别和促进现有系统的使用方面的进展受到阻碍。在这项提议中特别感兴趣的是结构多晶金属，在汽车、飞机和能源行业中具有核心重要性，在这些行业中，疲劳和断裂过程构成了巨大的建模和计算挑战。为了解决这些问题，动态高能X射线衍射(HEXD)实验最近已经上线，能够实时探测这些材料样品在加工或使用条件下的内部演变。由此产生的数据集既大(单个实验高达10TB)又复杂，因此使它们在材料设计过程中的分析和集成变得复杂。即使有广泛的人类交互，最先进的计算工具也只能提取这些数据集中包含的完整信息的一小部分。要实现这些数据和模型提供的潜力，需要全新的大数据类型的计算方法。本项目的工作就是开发这样一个工具集。在这个项目中特别关注的是复杂的、概率的、视频处理方法的使用和扩展，作为解决动态HEXD数据分析中的紧迫问题的基础。多晶样品的X射线衍射物理产生了由三维数据空间中的局域结构的时间演变集合组成的数据集，这些局部结构被称为“斑点”。将这些数据与计算塑性代码结合使用，需要将这些斑点与多晶体中的单个颗粒相关联，并跟踪这些随时间演变的结构集。到目前为止，解决这一索引问题的唯一工具本质上是静态的，最适用于材料样本处于原始状态的情况。这种动态索引问题与识别和跟踪视频场景中移动的对象的问题之间的相似之处促使PI团队开发的多假设跟踪方法适合并进一步开发用于分析HEXD数据。该方法基于在一大组假设上构建条件随机场，捕捉其中一个点可以与一个点相关联的方式。最优航迹和关联的估计是使用高效的图割方法进行的，使得整体方法非常适合于近实时实现。这一领域的现有工作将通过基于现有塑性代码为与点相关的特征(例如，质心位置、低阶矩)的演变构建动态模型并将这些模型结合到随机场中来扩展，以实现动态HEXD数据的多模型、多假设跟踪方法。

The past two decades have witnessed the development of accurate and efficient computational methods for a wide range of physical processes and the transition of these models into regularly used tools for product design and development within all sectors of the US economy. One important exception to this trend is in the field of material science where progress in creating new classes of materials and advancing the use of existing systems is hampered by the lack of validated computational models. Of particular interest in this proposal are structural polycrystalline metals, of central importance in the automotive, aircraft, and energy industries, where the processes of fatigue and fracture pose significant modeling and computational challenges. To resolve these issues, dynamic high-energy X-ray diffraction (HEXD) experiments have recently come on line that are capable of probing the internal evolution of samples of these materials in real time as they are subject to processing or service conditions. The resulting data sets are both large (up to 10Tb for a single experiment) and complex thereby complicating their analysis and integration within the material design process. Even with extensive human interaction, state-of-the-art computational tools can extract only a tiny fraction of the full information contained in these data sets. Realizing the potential offered by these data and models requires fundamentally new Big Data-type of computational methods. The work in this project is aimed at developing such a tool set. Of specific concern in this project are the use and extension of sophisticated, probabilistic, video processing methods as the basis for addressing a pressing problem in the analysis of dynamic HEXD data. The physics of X-ray diffraction from polycrystalline samples gives rise to data sets comprised of temporally evolving collections of localized structures, referred to as "spots," in a three-dimensional data space. Use of these data in conjunction with computational plasticity codes requires that these spots be associated with individual grains in the polycrystal and that these sets of evolving structures be tracked over time. To date, the only tools for addressing this indexing problem are static in nature and function best for cases where the material sample is in a pristine state. Similarities between this dynamic indexing problem and the problem of identifying and tracking objects moving in a video scene motivate the adaptation and further development of a multi-hypothesis tracking approach developed by the PI team to the analysis of HEXD data. The method is based on the construction of a conditional random field over a large set of hypotheses capturing ways in which spots can be associated with one. Estimation of the optimal tracks and association is carried out using efficient graph cut methods making the overall approach well suited to near real time implementation. The existing work in this field will be extended through the construction of dynamic models for the evolution of features associated with the spots (e.g., centroid location, low order moments) based on existing plasticity codes and incorporation of these models into the random field to achieve a multiple model, multi-hypothesis tracking approach for dynamic HEXD data.