From Limited Data to the Deformation Field in Metals: A Machine Learning Driven Approach

在材料发生灾难性故障之前发出警告可以挽救生命并降低成本。在失效之前，材料可能会经历内部变化，从而产生高频应力波，称为声发射。金属中声发射的测量为量化缺陷及其运动提供了一种独特的方法。然而，解释声发射是一个长期存在的挑战。这项研究将结合实验和机器学习工具，通过识别和解码每种变形机制的不同声发射特征来应对这一挑战。这项研究将产生一种独特的方法，为实验提供一个窗口，了解目前仅通过表面测量无法获得的基本变形机制。这将使人们能够对金属在变形过程中的行为有新的基本了解。这项研究还与教育和外联工作相结合。这项工作的结果将被整合到关于固体力学和材料工程的机器学习的新课程中。来自服务不足/代表性不足群体的马里兰州高中生也将参与研究实习机会。我们将利用基于物理的建模、机器学习和实验相结合的方法：(1)开发声发射实验的“数字孪生”，以正向预测单晶Ni微柱变形过程中与复杂滑移雪崩相关的声发射面波；(2)明确评估/仔细审查文献中现有的唯象声发射模型，并开发新的基于物理的理论模型，以确定位错塑性与声发射信号之间的相互联系；(3)根据表面声发射测量结果预测实验观察到的三维体积中真实的滑移局部化；(4)训练深度算子网络(DeepONets)用于声发射的正向预测和潜在变形机制的反向预测；以及(5)通过对单晶Ni微晶体的原位扫描电子显微镜微压缩实验和声发射测量来验证正向和反向预测。为了结束开发的模型和实验之间的循环，我们还将利用经过培训的DeepONet的实验结果来基本了解微压缩实验中位错雪崩的潜在变形机制，目前仅基于表面测量和载荷-位移测量很难解释这些机制。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Having warning before a catastrophic failure of a material occurs can save lives and reduce costs. Prior to failure a material may undergo internal changes which generate high-frequency stress waves, referred to as acoustic emissions. Measurements of acoustic emissions in metals provide a unique approach for quantifying defects and their movements. However, interpreting acoustic emissions is a longstanding challenge. This research will address this challenge by identifying and decoding the distinct acoustic emission signatures of each deformation mechanism with a combination of experiments and machine learning tools. This research will result in a unique method that endows experiments with a window into the fundamental deformation mechanisms that are not currently accessible from surface measurements alone. This will enable new basic knowledge of the behavior of metals during deformation. This research is also integrated with education and outreach. Results from this work will be integrated into a new course on machine learning for solid mechanics and materials engineering. Maryland high-school students from under-served/under-represented groups will also be engaged in research internship opportunities. We will utilize integrated physics-based modeling, machine learning, and experiments to: (1) develop a “digital twin’’ of acoustic emission experiments to forward predict the acoustic emission surface waves associated with complex slip avalanches during the deformation of single crystal Ni micropillars; (2) definitively assess/scrutinize existing phenomenological acoustic emission models in literature, and develop new physics-based theoretical models that identify the interconnections between dislocation-based plasticity and acoustic emission signals; (3) predict the true experimentally observed slip localization in the 3D volume from the surface acoustic emission measurements; (4) train deep operator networks (DeepONets) for forward predictions of acoustic emission and inverse predictions of the underlying deformation mechanisms; and (5) validate the forward and inverse predictions through coupled in situ scanning electron microscopy microcompression experiments and acoustic emission measurements on single-crystal Ni microcrystals. To close the loop between the developed models and the experiments, we will also utilize the trained DeepONets on the experimental results to gain fundamental understanding of the underlying deformation mechanisms during dislocation avalanches in micro-compression experiments, which are currently difficult to interpret based on surface measurements and load-displacement measurements alone.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.