EAGER: Reconstruction and Optimal Design of Multi-scale Material Systems through Deep Networks

这一早期概念探索性研究(AGUGER)赠款支持基础研究，以开发可扩展的计算设计工具，以实现高效和有效的材料设计。计算材料设计(CMD)，如确定最佳的材料微结构以实现所需的性能，受到越来越多的关注，因为复杂的材料设计随后可以利用先进的加工技术，如添加剂制造来实现。从概念上讲，解决CMD问题涉及到在问题空间中迭代搜索最佳解。由于寻找解决方案的成本对空间的大小很敏感，缺乏成本效益阻碍了现有CMD方法在复杂材料系统中的应用，其中材料设计的好坏取决于多个长度尺度上的微观结构的许多细节。CMD工具的使用将使发现关键微结构模式和降低问题空间的维度成为可能。该研究将为具有优异耐久性和结构健康的高性能结构材料提供高效的微观结构设计和验证。因此，这项研究的结果将使美国的各个行业及其经济和社会受益。材料科学、工程设计、制造和数据科学所需的无缝集成将有助于扩大学生的参与度，并对工程教育产生积极影响。其中包括(1)生成性统计模型，它在多个长度尺度上学习重要的局部微结构模式，并提供微结构及其低维设计表示之间的双向细节保持转换；(2)使用基于物理的模拟和主动学习，为工艺-结构-属性映射提供经济高效的建模框架，以促进有效的工艺和微结构级别的解空间探索和优化设计。将解决关键挑战，通过考虑学习特征的物理可解释性来实现可伸缩微结构特征学习。为此，将研究新的深层网络结构和学习技术。工艺-结构映射的统计建模将通过基于物理的模拟工具来实现，其中明确地考虑了相形态和结晶学信息；结构-性能映射的建模将通过一种新颖的晶格-粒子模拟方法来实现。

This EArly-concept Grant for Exploratory Research (EAGER) grant supports fundamental research to develop scalable computational design tools to enable efficient and effective materials design. Computational material design (CMD), such as identifying optimal material microstructures to achieve desirable performance, receives a growing interest as sophisticated material designs can be subsequently realized using advanced processing techniques such as additive manufacturing. Conceptually, solving CMD problems involves iterative search for the best solutions in a problem space. Since the cost of solution searching is sensitive to the size of the space, the lack of cost-efficiency hampers the application of existing CMD approaches to complex material systems, where the goodness of the material design depends on numerous details of the microstructure on multiple length scales. The use of CMD tools will enable the discovery of critical microstructure patterns and the reduction of the dimensionality of the problem space. The research will lead to efficient microstructure design and validation for high performance structural materials with superior durability and structural health. Therefore, results from this research will benefit various U.S. industries, and its economy and society. The required seamless integration of material science, engineering design, manufacturing, and data science will help to broaden student participation and positively impact engineering education.Two core technical enablers for tractable CMD will be developed. These include (1) a generative statistical model that learns important local microstructure patterns at multiple length scales, and provides a two-way detail-preserving conversion between microstructures and their low-dimensional design representations; and (2) a cost-effective modeling framework for processing-structure-property mapping using physics-based simulations and active learning, to facilitate effective processing- and microstructure-level solution space exploration and optimal design. Key challenges will be addressed to enable scalable microstructure feature learning by taking into account the physical interpretability of the learned features. Novel deep network architectures and learning techniques will be investigated to this end. The statistical modeling of the processing-structure mapping will be achieved by a physics-based simulation tool where both phase morphology and crystallographic information are explicitly considered; the modeling of the structure-property mapping is achieved by a novel lattice-particle simulation method.