Collaborative Research: EAGER: ADAPT:Charting the Space of Material Microstructures with Artificial Intelligence

非技术总结材料科学的基本原理之一是材料的性质由结构决定。微结构或微米级(百万分之一米)的内部结构被特别确定为对建筑、制造和其他工业应用中使用的陶瓷和金属部件的机械强度、延展性和断裂韧性等物理性能至关重要。由于未来特殊材料的微观结构有可能甚至可能与传统材料不同，材料开发中的一个关键挑战是确定所有可行的微观结构。该奖项将支持采用数据科学和机器学习领域的领先方法来应对这一挑战的研究和教育活动。具体地说，这项研究将把关于微结构的物理意义比较的专家知识整合到机器学习模型中，为探索可能的微结构提供一种系统的方法，包括以前实现的和未实现的。这种方法还有望提高仅基于微观结构预测材料性能的模型的精度和效率。这一奖项将为数学和材料科学的本科生和研究生创造在不同学科和机构之间进行交叉培训的机会。数学学生将从与材料科学家的互动中受益，反之亦然。此外，PI将创建用户友好的软件，使建议的算法广泛适用于研究人员和工业从业者以及研究类似几何结构的其他学科的个人。技术总结该奖项支持微结构状态空间的新表示法的开发，该表示法需要保留足够的信息来预测材料的物理性质，并要求其足够低维和通用，以作为灵活材料数据库的基础。计算材料设计的概念依赖于这样的基本思想：(I)微结构可以被表示为适当状态空间中的点，(Ii)该状态空间指定了足够的信息来准确地预测材料性能，以及(Iii)可以使用优化例程在状态空间中搜索具有所需性能的微结构。在这个研究计划中，PI将采用数据科学和机器学习中的领先方法，以发现适用于各种材料类别的这种微结构空间的实用表示。在形式上，将通过三个目标来提高实验微结构数据的特征提取、分类和可解释性。目的一：定义和实施物理激励的度量标准，以评估局部和全球尺度上微结构的相似性。目的II：利用具有流形学习的局部度量来构造窗口空间的坐标表示，并将这些坐标与新的机器学习技术相结合来预测材料属性。目标三：学习窗口分布空间的坐标表示，并利用它来构建概念验证微观结构数据库。该奖项将为数学和材料科学的本科生和研究生创造在不同学科和机构之间进行交叉培训的机会。数学学生将从与材料科学家的互动中受益，反之亦然。此外，PIS将创建用户友好的软件，使建议的算法广泛适用于研究人员和工业从业者以及研究类似几何结构的其他学科的个人。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。

NONTECHNICAL SUMMARYOne of the fundamental principles of materials science is that material properties are determined by structure. The microstructure, or the internal structure at the micron scale (one millionth of a meter), is specifically identified as being essential to physical properties including the mechanical strength, ductility, and fracture toughness of ceramic and metal components used in construction, manufacturing, and other industrial applications. Since it is possible and even likely that microstructures of exceptional materials of the future will not resemble those of conventional materials, a key challenge in material development is the determination of the all feasible microstructures. This award will support research and education activities that will adapt leading methods in data science and machine learning to address this challenge. Specifically, the research will integrate expert knowledge about physically-meaningful comparisons of microstructures into machine learning models to provide a systematic method for exploring possible microstructures, both previously realized and unrealized ones. This approach is also expected to improve the accuracy and efficiency of models to predict material properties on the basis of microstructure alone. This award will create opportunities for undergraduate and graduate students in mathematics and materials science to be cross-trained between disciplines and institutions. The mathematics students will benefit from interactions with materials scientists and vice versa. In addition, the PIs will create user-friendly software to make the proposed algorithms widely accessible, both to researchers and industrial practitioners and to individuals in other disciplines studying structures with similar geometry.TECHNICAL SUMMARYThis award supports the development of a new representation of microstructure state space that balances the need to retain enough information to predict physical properties of materials with the requirement that it be sufficiently low-dimensional and general to serve as the basis for a flexible materials database. The concept of computational materials design relies on the underlying ideas that (i) a microstructure can be represented as a point in an appropriate state space, (ii) this state space specifies enough information to accurately predict material properties, and (iii) optimization routines could be used to search the state space for microstructures with desirable properties. In this research program, the PIs will adapt leading methods in data science and machine learning to discover a practicable representation of this microstructure space applicable to a variety of material classes. Formally, the feature extraction, classification, and interpretability of experimental microstructure data will be improved by achieving three aims. Aim I: Define and implement physically-motivated metrics to evaluate the similarity of microstructures on both local and global scales. Aim II: Leverage the local metric with manifold learning to construct a coordinate representation for the space of windows, and apply these coordinates in conjunction with new machine learning techniques to to predict material properties. Aim III: Learn a coordinate representation for the space of window distributions and use it to construct a proof-of-concept microstructure database.This award will create opportunities for undergraduate and graduate students in mathematics and materials science to be cross-trained between disciplines and institutions. The mathematics students will benefit from interactions with materials scientists and vice versa. In addition, the PIs will create user-friendly software to make the proposed algorithms widely accessible, both to researchers and industrial practitioners and to individuals in other disciplines studying structures with similar geometry.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.