New approaches for a reliable fatigue life prediction of powder metallurgy components using machine learning

Abstract

Up to now, no consistent fatigue assessment approach of powder metallurgy (PM) components is available. For some materials and for some parameters, such as the density, a relationship to the fatigue strength is known; however, for other materials, such relationships are unknown. Based on an extensive data set with 828 test series, the present work addresses this problem by conceiving and applying five machine learning (ML)-based approaches to increase the accuracy of the prediction of the fatigue life as well as to predict the scatter of unknown data as precisely as possible. With the elaborated procedure, on the one hand, a scatter range of T S ¼ 1 : 1:40 can be achieved on completely unknown data. On the other hand, by using a newly defined loss function, the standard deviation of unknown data can be predicted very accurately. The findings provide the basis for further research on cost and efficiency optimized design of PM components through better estimation of fatigue life.

Fatigue crack growth simulation by extended finite element method: A review of case studies

“…Since all of them are based on experimental or analytical data, they are not presented here, but some of them are referenced here for the benefit of readers. [136][137][138][139][140][141][142] In addition, many other recent papers presented also this important topic, some of them described briefly in the following text.…”

“…Excellent source for such an overview is provided by the special issue published couple of months ago in Fatigue & Fracture of Engineering Materials & Structures under the title “Data science and machine learning for fatigue and fracture assessment.” 135 As presented and explained in Editorial, 11 papers described the use of data sciences and ML in structural durability investigations with a particular emphasis on material cyclic behavior and fractures. Since all of them are based on experimental or analytical data, they are not presented here, but some of them are referenced here for the benefit of readers 136–142 …”

Machine learning in materials research: developments over the last decade and challenges for the future

“…First, scikit-learn implements a variety of techniques that can be useful for many different types of materials machine learning. For example, it can be applied to predict the band gaps of solids 19 , to predict the strength of cement composites 20 , to associate processing conditions with final properties in batteries 21 , to predict the fatigue life of powder metallurgy components 22 , or for many other materials tasks. Furthermore, although scikit-learn is missing the capability to implement more complex deep learning models, the small data set sizes of many materials problems often make it practical to use more conventional machine learning algorithms that have fewer parameters to train.…”

Estimation of Mass and Lengths of Sintered Workpieces Using Machine Learning Models

“…Few research papers have demonstrated the suitability of ML models in the quality assessment of sintered components. These models have been employed for the estimation of mechanical and fatigue properties [22], [24], or density estimation of sintered bronze [23]. Here, we focus on the prediction of different QCs, i.e., mass and lengths, for more complex, i.e., multilevel, workpieces.…”