EAGER: Type I: Liquid metal embrittlement of engineering alloys by eutectic gallium indium: Data-driven experimental design using sequential learning

该奖项支持在MATDAT18数据马拉松活动上开展的一项新合作的研究和教育，重点是利用数据科学的方法在材料科学中的挑战性问题上取得进展，在液态金属脆化机制方面。当某些液态金属与特定的固体金属接触时，固体金属的强度和/或延展性会发生灾难性的降低；这被称为“液态金属脆化”。虽然对液态金属脆化的研究已经有一个多世纪了，但对这一现象还缺乏充分的了解。目前还没有办法预测在给定条件下液态金属脆化的发生或严重程度。pi的目标是使用一种方法，计算机可以从许多研究中获得的数据中“学习”，以创建一个模型，该模型可以预测液态金属脆化的严重程度，作为实验条件的函数，包括液体成分，固体成分，温度，变形率和固体金属的微观结构，这些都是通过强大的光学显微镜可见的。作为本研究的一部分创建的机器学习模型可以使液态金属在工程应用中使用，例如在可拉伸电路中，并允许未来研究液态金属脆化的基本物理机制。该项目非常重视各年龄段学生的教育和专业发展。具有材料科学背景的研究生和本科生研究人员将接受传统实验室技能和工程师数据科学方法的培训。此外，针对初高中学生的拓展模块将作为该项目的一部分开发，并通过工程项目中的少数民族和北卡罗莱纳州立大学的工程场所等项目分发给更广泛的社区。该奖项支持在MATDAT18数据马拉松活动上开展的一项新合作的研究和教育，重点是利用数据科学的方法在材料科学中的挑战性问题上取得进展，在液态金属脆化机制方面。迄今为止，还没有成功地开发出液态金属脆化的预测现象学或力学模型。液态金属的脆化现象非常复杂，其脆化行为几乎取决于所测试的每一个实验变量，包括温度、应变速率、固体金属晶粒尺寸、固体成分、液体成分等等。由于这种复杂的现象学和独立变化大量涉及的实验挑战，纯经验的液态金属脆化研究是棘手的。本项目采用另一种方法来建立预测液态金属脆化模型：Citrination平台将进行顺序学习。在这种方法中，使用初步数据训练初始模型，并用于建议下一轮最有可能提高模型预测能力的实验。PI最近根据从文献中提取的数据开发了一个初始模型。该模型用于建议初步实验，这些实验被进行并用于完善模型。然而，需要进一步的迭代来实现模型的预测能力。pi的目标是迭代地改进模型。一旦模型具有足够的预测能力，这项工作的第二个目标是测试液态金属脆化不是一个整体现象，而是由几个不同的机制组成的假设。这项工作的最后一个主要目标是为每个确定的潜在机制和未来机制研究的理想候选者确定“原型”系统。如果多种机制的假设是支持的，这可以调和似乎矛盾的报告的液态金属脆化行为目前在文献中。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。

NONTECHNICAL SUMMARYThis EAGER award supports research and education involving a new collaboration kindled at the MATDAT18 Datathon event focused on using the methods of data science to make progress on challenging problems in materials science, in the mechanisms of case liquid metal embrittlement. When certain liquid metals come into contact with specific solid metals, the solid metals can undergo a catastrophic reduction in strength and/or ductility; this is termed "liquid metal embrittlement." While liquid metal embrittlement has been studied for over a century, a full understanding of the phenomenon is lacking. There is currently no means to predict the occurrence or severity of liquid metal embrittlement under given conditions. The PIs aim to use a method where computers can "learn" from the data obtained from many studies to create a model that can predict the severity of liquid metal embrittlement as a function of the experimental conditions, including liquid composition, solid composition, temperature, deformation rate, and microscopic structure of the solid metal which is largely visible through powerful optical microscopes. The machine learning model created as part of this research may enable the use of liquid metals in engineering applications, such as in stretchable circuits, and allow for future study of the fundamental physical mechanisms responsible for liquid metal embrittlement. The project strongly emphasizes the education and professional development of students of all ages. Graduate and undergraduate researchers with a materials science background will be trained in both conventional laboratory skills and in data science methods for engineers. Additionally, outreach modules targeted to middle and high school students will be developed as part of this project and distributed to the broader community through programs including the Minorities in Engineering Program and The Engineering Place at North Carolina State University. TECHNICAL SUMMARYThis EAGER award supports research and education involving a new collaboration kindled at the MATDAT18 Datathon event focused on using the methods of data science to make progress on challenging problems in materials science, in the mechanisms of case liquid metal embrittlement. To date, no predictive phenomenological or mechanistic models of liquid metal embrittlement have been successfully developed. The phenomenon of liquid metal embrittlement is incredibly complex, with embrittlement behavior shown to depend on nearly every experimental variable ever tested including temperature, strain rate, solid metal grain size, solid composition, liquid composition, and more. Due to this complex phenomenology and the experimental challenge of independently varying the large number of involved, purely empirical studies of liquid metal embrittlement are intractable. This project takes an alternative approach to establish a predictive liquid metal embrittlement model: the Citrination platform will be used to conduct sequential learning. In this approach, an initial model is trained using preliminary data and used to suggest the next round of experiments which will have the greatest likelihood of improving the predictive capability of the model. The PI recently developed an initial model trained on data extracted from the literature. The model was used to suggest preliminary experiments, which were conducted and used to refine the model. However, further iterations are required for the model to achieve predictive capability. The PIs aim to iteratively refine the model. Once the model has sufficient predictive capability, a second objective of this work is to test the hypothesis that liquid metal embrittlement is not a monolithic phenomenon but is composed of several distinct mechanisms. The last main objective of this work is to identify "archetypal" systems for each identified potential mechanism and ideal candidates for future mechanistic study. If the hypothesis of multiple mechanisms is supported, this could reconcile seemingly contradictory reports of liquid metal embrittlement behavior present in the literature.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.