Multi-Information Source Value of Information Based Design of Multiphase Structural Materials

本研究项目的目标是为复杂、多部件、多阶段的先进钢的设计创建一个基于信息价值的多信息源框架。在材料科学中，利用化学-加工-微观结构-性能范式的材料设计问题通常是昂贵的，耗时的，并且受可用资源的限制。因此，为了实现设计目标，必须仔细选择要查询的信息源，它可以是一个实验，一个计算模拟，专家意见的启发等，基于对设计问题的附加价值，以及与执行查询相关的时间/成本。与传统钢相比，先进钢中复杂的微观结构显著提高了强度和延展性。这使得这种钢材对汽车应用特别有吸引力，因为汽车重量减轻和耐撞性是至关重要的。这种方法是一种全新的、具有潜在变革意义的看待材料设计的方式。在这里，实验、模拟和先验知识在同一水平上被考虑，潜在地减轻了已建立范式的主要缺点，因为不再需要在计算或实验方法的输入和输出之间进行任何对齐。该设计框架隐含地考虑了信息源的保真度，能够有效地将有限的资源分配到信息经济意义上的材料设计任务中。该研究还将与美国国家科学基金会在德克萨斯农工大学的研究和培训计划相结合，并通过德克萨斯农工大学现有的项目扩展到更广泛的社区。本研究的目的是在一个促进多种信息源集成的框架内，利用材料设计中的化学-加工-微观结构-性能范式。它规定了平衡探索和利用设计空间的最佳信息获取协议。该框架将扩展最先进的知识梯度策略，以实现基于信息源的信息价值优化查询。该框架的主要贡献包括：通过拟合高斯过程模型来实现对相关信息源的并行查询；利用高维模型表示将信息源与不同的输入/输出空间结合起来；以及多目标策略，通过将先前查询的信息重新加权为最先进的超容量指标，提供预期改进的标量指标。与许多现有的材料设计框架相反，该框架直接将材料的化学/加工与性能联系起来，避免了最优性和可行性之间的差异问题。该框架将能够适应不同分辨率模型的力学性能预测，更重要的是，允许在同一统一框架内整合实验和先验知识。虽然演示的重点是低合金多相高级钢，但该方法可能适用于广泛的材料系统。

The goal of this research project is to create a multi-information source value-of-information based framework for the design of complex, multi-component, multi-phase advanced steels. In materials science a material design problem that exploits the chemistry-processing-microstructure-properties paradigm is often expensive, time consuming and limited by the available resources. Thus, to achieve a design goal, one must carefully select the information source to query, which could be an experiment, a computational simulation, elicitation of expert opinion, etc., based on the value added to the design problem, as well as the time/cost associated with performing the query. The complex microstructures in advanced steels result in significant improvements in strength and ductility compared to conventional steels. This makes such steels especially attractive for automotive applications, where vehicle weight reduction and crashworthiness are paramount. The approach is a completely new and potentially transformative way of looking at materials design in general. Here, experiments, simulations, and prior knowledge are accounted for at the same level, potentially alleviating major drawbacks of established paradigms in that there is no longer any required alignment between inputs and outputs of computational or experimental methods. The information source fidelity is implicitly accounted for, and the design framework can efficiently allocate limited resources to the materials design task in an information-economic sense. The research will also be integrated with the NSF Research and Traineeship Program at Texas A&M and outreach to broader communities through existing programs at Texas A&M.The objective of this research is to exploit the chemistry-processing-microstructure-properties paradigm in materials design within a framework that facilitates the integration of multiple information sources. It prescribes optimal information acquisition protocols that balance exploration and exploitation of the design space. This framework will extend state-of-the-art knowledge gradient policies for value of information based optimal querying of information sources. Key contributions of the framework include enabling parallel querying of correlated information sources by fitting Gaussian process models to learned correlations; incorporation of information sources with differing input/output spaces by leveraging high dimensional model representations; and multi-objective policies that provide scalar indicators of expected improvement through novel importance reweighting of prior queried information into state-of-the-art hypervolume indicators. Contrary to many existing materials design frameworks, the framework directly connects materials chemistry/processing to property, avoiding issues related to discrepancies between optimality and feasibility. The framework will be able to accommodate predictions of mechanical properties with models of different resolution and, more importantly, allows for the integration of experiments and prior knowledge within the same unified framework. Although the demonstration focus is on low alloy multi-phase advanced steels, the methodology is potentially applicable to a wide range of materials systems.