QRM: Hybrid Adversarial-Training Methods for 3D Virtual Microstructures

该基金将支持开发严格的科学方法来量化工程材料的内部结构或微观结构的研究，推进科学前沿并使美国经济受益。微观结构决定了工程材料的性质和性能，但通常用定性而不是定量的术语来描述。该奖项支持基础研究，以建立量化微观结构的科学工具，并使微观结构的数字表示能够用于计算模拟，以预测结构和功能应用中的材料性能。该项目使用最先进的数据科学和机器学习（ML）来构建定量准确、高保真的虚拟微结构，并将这些工具与基于物理的计算模型相结合，以预测服役中的材料如何应对极端环境。由此产生的能力允许为制造业、基础设施和航空航天设计高性能材料，并有可能改变材料从实验室到商业应用的方式。该项目使用最先进的数据科学和机器学习来构建具有有限不确定性的虚拟微观结构。该方法建立在DREAM.3D合成微观结构软件中存在的“形状优化器”和生成对抗网络（GAN）的新生能力之上。这些方法将在统计上与实验数据相结合，以获得定量表示的准确性。相场模拟中的虚拟微结构将作为机器学习预测的一个质量指标。混合方法探索和利用GAN方法来抵消形状优化器和相变方法中的弱点。目标是获得完整、真实、准确的虚拟结构。为了展示新混合方法的收益，不确定性量化（UQ）技术将同时开发，使用虚拟“材料幻影”，以及来自合成结构和实验的材料描述符的评估。这项工作基于材料工程和机器学习领域的开源软件和公共数据集。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

This grant will support research that will develop a rigorous scientific method in quantifying the internal structure, or microstructure, of engineering materials, advancing the frontiers of science and benefiting the US economy. Microstructure determines the properties and performance of engineering materials, but is often described in qualitative rather than quantitative terms. This award supports fundamental research to build the scientific tools to quantify microstructure, and enable the creation of digital representations of microstructure that can be used for computational simulations to predict material performance in structural and functional applications. This project uses state-of-the-art data science and machine-learning (ML), to build quantitatively accurate, high-fidelity virtual microstructures, and combines these tools with physics-based computational models to predict how materials in service respond to extreme environments. The resulting capabilities allow for the design of high-performance materials for manufacturing, infrastructure, and aerospace, and have the potential to transform how materials are brought from the laboratory to commercial applications.This project uses state-of-the-art data science and machine-learning to build virtual microstructures with bounded uncertainty. The approach builds upon "shape-optimizers" existing within the DREAM.3D synthetic microstructure software, and the nascent capabilities of generative adversarial networks (GAN). These methods will be statistically coupled to experimental data for gains in quantitative representation accuracy. The virtual microstructures represented in phase-field simulations will serve as one quality metric for the ML prediction. The hybrid method explores and exploits use of the GAN methods to offset weaknesses within shape-optimizer and phase-transformation methods. The goal is to obtain complete, realistic, and accurate virtual structures. To demonstrate the gains from the new hybrid method, uncertainty quantification (UQ) techniques will be concurrently developed using both virtual "materials phantoms," and evaluations of material descriptors from both synthetic structures and experiments. The effort builds from open-source software and public datasets in the Materials Engineering and Machine Learning fields.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.