User login

Navigation

You are here

Discussion of fracture paper #32 Fatigue and machine-learning

esis's picture

The paper, "A machine-learning fatigue life prediction approach of additively manufactured metals" by Hongyixi Bao, Shengchuan Wu, Zhengkai Wu, Guozheng Kang, Xin Peng, Philip J. Withers in Engineering Fracture Mechanics 242 (2021) 107508, p. 1-10., adopts a very interesting view of the correlation between fault geometry and fatigue properties. A simplified statistical description of irregular faults in large numbers is used. The variety of fault shapes that appear during the production of 3D objects from powder metal is described in terms of the distribution of size, volume, and position.

The studied test specimens are produced by selective melting during the build-up of a powder bed of a granulated titanium alloy. Each new layer is fused together with the underlying solidified material. The heat is introduced by a focused high-energy ultraviolet light beam. An almost inevitable problem is small defects, typically of grain size. Naturally, the strength of the structure and especially the fatigue properties take a beating. The authors examine the defects using synchrotron X-ray tomography. After fatigue experiments, the results are used for a machine learning method based on extended linear regression.

The statistical description based on a few geometric and morphology parameters if of course better than the size of a hypothetical crack that we often use for fracture mechanical analyses. The correlation of the more realistic geometrical description with the fatigue limits swallows the entire series of events from fault, fatigue crack initiation, and growth to final rupture.

I guess it could be interesting to benchmark test could be to use available analytical solutions of interacting, cracks, holes, spheres, edges, etc. If necessary numerical ditos could be used. Stress intensity factors for cracks and stress criteria for other faults with smooth boundaries.  

The paper is nicely written and offers very interesting reading. To me, the paper also calls for a reflection. Very few scientific studies combine basic science from different disciplines and create something directly industrially useful. This present paper is a good example of that. 

For industrial applications perhaps the Kalman filter could lead to a speedier optimization since it recursively adds adjustment of the previous result after each new mechanical test. In terms of calculation, it is advantageous because it does not require recalculation after each new test. The process provides a good overview and the series of tests can be interrupted as soon as an appropriate convergence rate-based criterion is met.

It would be interesting to hear from the authors or anyone else who would like to discuss or provide a comment or a thought, regarding the paper, the method, or anything related.

Per Ståhle

Subscribe to Comments for "Discussion of fracture paper #32 Fatigue and machine-learning"

Recent comments

More comments

Syndicate

Subscribe to Syndicate