Taylored Material Properties via Microstructure Optimization: Machine Learning for Modelling and Inversion of Structure-Property-Relationships and their Application to Sheet Metals

生产具有特定、期望性能的材料的加工控制动作的推导是“过程控制”-“微观结构实例化”-“材料性能”因果链的“逆问题”。该项目的主要目标是通过使用机器学习领域的现代方法，为解决这个问题创造一个新的基础。该反演将由两个明确分离的部分组成：“属性-结构-映射”和“结构导向的最优过程控制”。该项目的重点在于研究和开发方法，这些方法可以反转材料的结构-性能-关系，这与工业相关。这种反演是微结构设计和相关生产过程优化控制的基础。另一个目标是发展最优控制方法，产生具有期望性质的结构。开发的通用方法如何用于解决具有专用性能的材料生产的具体逆问题，将通过将它们应用于钣金生产过程来展示。核心目标包括开发与技术相关的“结构-属性-映射”的反转方法，以及通过监督和无监督机器学习实现高效微观结构表示的方法。基于强化学习的自适应加工路径优化方法将用于制造过程的自适应最优控制。我们期望这项工作的结果将导致对材料的工艺-结构-性能-关系的技术相关的越来越深入的了解。该项目产生的仪器还将促进新材料和工艺控制的经济有效发展。

The derivation of processing control actions for the production of materials with certain, desired properties is the "inverse problem" of the causal chain "process control" - "microstructure instantiation" - "material properties". The main goal of the proposed project is the creation of a new basis for the solution of this problem by using modern approaches from the field of Machine Learning. The inversion will be composed of two explicitely separated parts: "Property-Structure-Mapping" and "Structure-guided Optimal Process Control".The focus of the project lies on the investigation and development of methods which allow an inversion of the structure-property-relations of materials, which are relevant in the industry. This inversion is the basis for the design of microstructures and for the optimal control of the related production processes. Another goal is the development of optimal control methods yielding exactly those structures which have the desired properties. How the developed generic methods are used to solve a concrete inverse problem of the production of a material with dedicated properties, will be shown by applying them to sheet metal production processes. The core goals include the development of methods for inverting technologically relevant "Structure-Property-Mappings" and methods for efficient microstructure representation by supervised and unsupervised machine learning. Adaptive processing-path-optimization methods, based on reinforcement learning, will be developed for adaptive optimal control of manufacturing processes. We expect that the results of this work will lead to an increasing insight into technological relevant process-structure-property-relationships of materials. The instruments resulting from the project will also promote the economically efficient development of new materials and process controls.