Solidification of the weld pool in ferritic low-alloyed and C-Mn steels leads to a variety of microstructures. They play a key role in the mechanical properties and mainly depend on the chemical composition, welding process, and welding parameters [1]. Hence, reliable identification and quantification of those microstructures is required to predict the behaviour of the weld.

Those microstructures are the result of the decomposition of austenite in different ferritic products. In the literature a number of microstructure nomenclatures can be found [2]. Here we defined four principal microstructures that are of interest in the composition range under study (low-alloyed and C-Mn steel): primary ferrite (PF), Widmanstätten ferrite (WF), acicular ferrite (AF) and microphases.

Because those microstructures are ferritic products, under the microscope it is often difficult to make a clear distinction between the different phases, and only an experienced eye can conduct a reliable quantification using metallographic methods. Classic methods of image filtering often fail to differentiate the phases that are present. A neural network approach could be an alternative method to conduct the microstructural quantification, especially for semantic segmentation using optical microscopy after etching. A classic architecture for this type of applications is the convolutional neural network. Most of the work has been done four wrought steel and coupled optical images with scanning electron microscopy and EBSD [3,4].

In this work we present a supervised learning of a U-net for automatic semantic segmentation of high magnification optical microscopy after chemical etching (Nital 2%). Semantic segmentation is performed on four classes (PF, WF, AF and other microphases) as shown in Figure 1. The goal is to allow the quantification of microstructure on large area of the weld pool.

The metric used is a Dice coefficient and after a training of 120 epoch a mean score among the classes of 82% is achieved with a better score on AF and lower one on WF with confusion between WF and PF. Acicular ferrite is the most desired microstructure in welding steel due to its good properties in toughness [5], then the good segmentation by the model on this microstructure could be used for reliable prediction of mechanical performance.

References

[1] H.K.D.H Bhadeshia, L.-E. Svensson; Mathematical Modelling of Weld Phenomena, eds H. Cerjak, K.E. Easterling, Institute of Materials, London, 1993, 109-182.

[2] G. Thewlis; Materials Science and Technology, 2004, 20:2, 143-160.

[3] T. Martinez Ostormujof, R.R.P. Purushottam Raj Purohit, S. Breumier, N. Grey, M. Salib, L. Germain; Materials Characterization, 2022, vol. 184, 111638.

[4] S.M. Azimi, D. Britz, M. Engstler, et al; Sci Rep, 2018, vol. 8, 2128.

[5] D.J. Abson; Science and technology of welding and joining, 2018, vol. 23, no. 8, 635-648.