SEM images data set development of ML models to predict hydrogen embrittlement in welded 316 L/304 L stainless steels

• Materials, Manufacturing, and Hydrogen Exposure : Gas Metal Arc Welding (GMAW) was performed on 316 L/304 L stainless steel samples with varying weld gaps and welding parameters. The samples were exposed to hydrogen gas at 80 bar for 150 h and analyzed for surface defects using SEM. • Analysis of Microstructure through EBSD Studies : EBSD analysis revealed significant phase transformations and grain boundary modifications in 304 L and 316 L stainless steels due to hydrogen exposure. Austenite content increased while cementite decreased, indicating hydrogen’s impact on microstructural stability. • SEM Image Processing for ML Training : SEM images were curated and pre-processed for training a CNN model using AlexNet to identify hydrogen embrittlement defects. Data augmentation techniques improved the model’s robustness, achieving 90% accuracy in predicting defects on each segregated image. • Preliminary Revelation of Results : Initial SEM analysis showed significant hydrogen-induced defects in welded zones and heat-affected zones. The CNN model demonstrated high accuracy in predicting defects, but further validation with a comprehensive dataset is recommended. • Ongoing Research and Project Funding : This work is part of the British Council-funded UK India Education and Research Initiative (UKIERI) project (Ref. UKIERI / SPARC / 01 / 06). The research is ongoing, and further studies are being conducted on various bars along with major ML/AI development frameworks. This study focuses on the machine learning (ML)-based identification of defects caused by hydrogen embrittlement (H2E) in the welded zones of 316 L/304 L stainless steels. It involves developing a robust SEM image dataset to train ML models for accurate defect identification. Initially, Gas Metal Arc Welding (GMAW) samples were manufactured with weld gap variations of 0.8 mm, 1.2 mm, and 1.5 mm. The welding parameters used were: (i) welding speeds of 15 mm/sec and 10 mm/sec, (ii) wire feed rate of 5.5 m/min, and (iii) voltage of 15.5 V. The samples were then exposed to a hydrogen gas environment at a pressure of 80 bar for 150 h. When analyzed using scanning electron microscopy (SEM) & electron backscatter diffraction (EBSD), H2E was observed on the surfaces of the welded zones (WZ) and heat-affected zones (HAZ). These defects, validated through literature, were segregated / sectioned as defect-based feature images and stored as a dataset. A preliminary analysis of the images validated after with 16 DOE’s using AlexNet, a convolutional neural network (CNN)-based ML model, showed significant identification of these defects with 90 % accuracy. The trained models helped identify areas and understand previously unidentified defects. Through a focused discussion on defect detection, supported by validation using classification (CNN Accuracy, Precision, Recall, and F1-Score) and regression metrics (R2 and Success Rate), the article demonstrates the potential of ML-based approaches in advancing welding diagnostics.