Accurate fatigue life prediction of additive manufactured parts is critical for the reliability and safety assessment of the designs made for aerospace applications. The fatigue life depends on the cyclic stress experienced due to loads in operation, surface roughness, internal microstructure, and defects in the parts. The microstructure of a material contains signatures of the manufacturing process and post-processing experienced by the part. Incorporating microstructure information in fatigue life prediction is difficult using analytical and empirical relations. A data-driven machine learning framework can be used to model complex phenomena without solving the detailed underlying physics. Manual selection of important features from microstructure may not capture all the properties that affect fatigue. In this work, the fatigue data of Ti-6Al-4V alloy is collected from several sources and machine learning models are trained using surface roughness, stress cycles and microstructure images. The effect of utilizing microstructure images and their 2-point statistics data with convolutional neural networks and Gaussian process regression for prediction of fatigue life are demonstrated. Various methods of image processing, data preparation, and modeling techniques are studied and outcomes are discussed.

中文翻译：

---

一种基于微观结构敏感机器学习的方法，用于预测增材制造零件的疲劳寿命


增材制造零件的准确疲劳寿命预测对于航空航天应用设计的可靠性和安全性评估至关重要。疲劳寿命取决于由于运行载荷、表面粗糙度、内部微观结构和零件缺陷而承受的循环应力。材料的微观结构包含零件所经历的制造过程和后处理的特征。使用分析和经验关系将微观结构信息纳入疲劳寿命预测是很困难的。数据驱动的机器学习框架可用于对复杂现象进行建模，而无需解决详细的基础物理问题。从微观结构中手动选择重要特征可能无法捕获影响疲劳的所有特性。在这项工作中，从多个来源收集了 Ti-6Al-4V 合金的疲劳数据，并使用表面粗糙度、应力循环和微观结构图像训练机器学习模型。证明了利用微观结构图像及其 2 点统计数据与卷积神经网络和高斯过程回归预测疲劳寿命的效果。研究了图像处理、数据准备和建模技术的各种方法，并讨论了结果。