CDS&E/Collaborative Research: Interpretable Machine Learning for Microstructure-Sensitive Fatigue Crack Initiation from Defects in Additive Manufactured Components

近几十年来，实验和计算方法的进步为许多工程和科学应用提供了丰富的数据。然而，这些数据并不容易转化为工程知识。该项目的目标是开发一种机器学习方法，以促进这种数据到知识的转换，以更好地理解材料的完整性。这些知识可以帮助理解增材制造部件中的机械行为，例如疲劳。 所开发的方法将改善传统的材料认证过程，这是昂贵的。该项目的结果可能会降低消费者成本并增加采用，最终可能会推动美国经济发展。为了吸引后代并促进包容性，K-12学生将在犹他州工程日和普渡太空日与用户友好的界面进行互动，以实践演示学习自然规律。提高将数据转化为知识的成功性和可靠性需要将重点转向可解释性和可解释性，以延续合理的科学和工程原则。为此，基于遗传编程的符号回归（GPSR）将被用来模拟结构材料的疲劳损伤。然后，GPSR模型将使用从材料模拟、实验中生成的数据集进行训练，并在现有知识的指导下发现新的潜在机制，即，知识研究任务将解决一个易于处理的方法来模拟疲劳寿命预测和替代目前的做法微观结构依赖的机制。 具体而言，GPSR模型将在高能X射线衍射显微镜和晶体塑性有限元模拟数据的组合上进行训练。生成的GPSR模型将是增材制造金属中孔隙诱导的、微观结构相关的疲劳裂纹萌生的物理正则化多尺度均匀化。该奖项反映了NSF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。

Advancements in experimental and computational methods in recent decades have enabled production of a wealth of data for many engineering and science applications. However, these data do not readily translate into engineering knowledge. The objective of this project is to develop a machine-learning approach to facilitate this data-to-knowledge translation to better understand materials integrity. Such knowledge can help understand mechanical behaviors, such as fatigue, in additive manufactured components. The developed approach will improve the conventional process of materials certification, which is prohibitively expensive. The outcome of this project could potentially reduce consumer costs and increase adoption, which may ultimately advance the U.S. economy. To engage future generations and promote inclusion, K-12 students will interact with a user-friendly interface for hands-on demonstration of learning natural laws at the Utah Engineering Day and Purdue Space Day.Increasing the success and reliability of translating data into knowledge requires a shifted focus toward explainability and interpretability to perpetuate sound science and engineering principles. To this end, Genetic Programming based Symbolic Regression (GPSR) will be utilized to model fatigue damage in structural materials. GPSR models will then be trained using generated data sets from materials simulations, experiments, and guided by existing knowledge to discover new underlying mechanisms, i.e., knowledge. The research tasks will address a tractable means to model microstructure-dependent mechanisms into fatigue life predictions and supplant current practices. Specifically, GPSR models will be trained on a combination of high-energy X-ray diffraction microscopy and crystal plasticity finite element simulation data. The generated GPSR models will be a physics-regularized multiscale homogenization of pore-induced, microstructure-dependent fatigue crack initiation in an additive manufactured metal.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.