Autonomous research for exploring structure-property linkages and optimizing microstructures

开发新材料在技术创新中发挥着重要作用。特别是工艺参数、局部材料结构和所得到的性能之间的关系引起了人们的极大兴趣。由于材料的局部结构本质上决定了其性质，因此结构-性质(SP)联系的产生和理解是至关重要的。设想的项目旨在快速和廉价地建立可靠的SP链接，从而确定具有最佳性能的现实的、可制造的材料结构。SP链接的试验性探索需要相当大的技术努力，这与高昂的成本相关。可实现的有限数据基础是现代机器学习方法应用的主要障碍。为此，基于模型的数值模拟可以以合成数据的形式提供补救措施，以减少复杂和昂贵的实验数量。此外，一个实用的相关算法需要自动生成合成微结构并对选定的样品进行数值模拟，以便系统地扩展实验数据库。该过程被称为自主研究，分四个步骤进行：1)开发具有丰富形态信息并允许快速重建的静止的、平移不变的微结构描述符。2)从这样的描述中有效地重建微结构是当前的研究课题。在初步结果的基础上，提出了一种具有普遍适用性的快速算法。3)确定有效性质的数值模拟通过模型和数值分辨率的层次化被有效地呈现，并辅以不确定性量化。4)结合计算的显微组织-性能对，建立了SP连杆机构，并对其最佳值进行了估计。基于这一估计，选择下一个样本并关闭自治环路。这四个步骤以与材料无关的方式开发和实施，并使用两个示例性微结构进行验证。作为SP联系和优化目标的有效性质同样具有一般性。在这个项目中，除了预测弹性性能外，损伤容限还被视为有效的临界能量释放率。基于模型的数值模拟和数据驱动算法的综合将有助于更深入地了解SP连接的作用机理，并加快材料的开发。

Developing novel materials plays an important role in technological innovations. Especially the relationship between the process parameters, the local material structure and the resulting properties is of great interest. Since the local structure of the materials essentially dictates its properties, the generation and understanding of structure-property (SP) linkages is of utmost importance. The envisioned project aims to establish reliable SP linkages fast and cheaply, thereby identifying realistic, manufacturable material structures with optimal properties.The experimental exploration of SP linkages requires considerable technological effort, which is associated with high costs. The achievable limited data basis is a major obstacle for the application of modern machine learning methods. To this end, model-based numerical simulations can provide a remedy in the form of synthetic data to reduce the number of complex and expensive experiments. Furthermore, a practically relevant algorithm requires the automated generation of synthetic microstructures and numerical simulations on selected samples in order to systematically extend the experimental data base. This procedure, referred to as autonomous research, is developed in four steps: 1) Stationary, translation-invariant microstructure descriptors are developed that are rich in morphological information and allow for a fast reconstruction. 2) Efficiently reconstructing a microstructure from such a description is a current research topic. Based on first preliminary results, a fast and universally applicable algorithm is developed. 3) The numerical simulations for the determination of effective properties are rendered efficiently through a hierarchy of models and numerical resolutions and are supplemented by uncertainty quantification. 4) Taking the previously calculated microstructure-property pairs calculated into account, an SP linkage is established and its optimum is estimated. Based on this estimate, the next sample is selected and the autonomous loop is closed.These four steps are developed and implemented in a material-independent manner and validated using two exemplary microstructures. The effective property as the objective of SP linkages and optimization is equally generic. In this project, beyond the prediction of elastic properties, damage tolerance is considered in terms of an effective critical energy release rate.Altogether, the synthesis of model-based numerical simulation and data-driven algorithms promises to provide a deeper understanding of the mechanisms of action underlying SP linkages, as well as to accelerate materials development.