A semantic segmentation algorithm for automated rapid melt pool identification from cross-sectional micrographs

Melt pool contours are an important empirical process measurement for metallic additive manufacturing. Melt pool boundaries indicate the maximal extent of the melting temperature contour. Melt pool boundaries characterize the effects of process parameters on melt pool size and variability and provide calibration and validation data for process simulations predicting thermal fields and microstructure. Serial cross sectioning of samples can produce numerous micrographs valuable for building robust melt pool contour datasets. However, the acquisition of contours from micrographs by hand is error prone and prohibitively time consuming for datasets large enough to capture variation within individual scan tracks and across a range of process conditions. In this work, we propose a semantic segmentation algorithm for automating melt pool contour acquisition from micrographs. The proposed algorithm is demonstrated on datasets from two different single scan track experiments conducted on a Powder Bed Fusion-Laser Beam platform using a bare Ti-6Al-4 V plate and an Inconel 718 plate with a 40 μm powder layer at different preheat temperatures. The algorithm was successfully trained and predicted the melt pool boundaries in the 300 image Ti-6Al-4 V dataset with 95% accuracy and an estimated reduction of 91% in processing time compared to data acquisition by hand. For the 1201 image Inconel 718 dataset, the algorithm achieved 89% accuracy and an estimated 99% time savings compared to hand measurements. The results for both datasets are demonstrated with a discussion about observed variability in the melt pool characteristics. •Semantic segmentation automates acquisition of melt pool contours produced by Powder Bed Fusion-Laser Beam from micrographs.•Datasets from 1501 Ti-6Al-4 V and Inconel 718 images are provided for melt pools at different preheat and process conditions.•The algorithm predicted melt pool boundaries with 95% accuracy (Ti-6Al-4 V dataset) and 89% accuracy (Inconel 718 dataset).