Transformer based spatially resolved prediction of mechanical properties in wire arc additive manufacturing

Metal additive manufacturing (MAM) provides remarkable design and component geometry freedom over various materials. One of the most recent MAM methods is the wire-arc additive manufacturing (WAAM) technique, which provides a higher deposition rate than other methods. This method also suffered from heterogeneity in location-based thermal profiles, leading to spatial variation in the properties of as-built mechanical properties, which become more complicated in the manufacturing design and process of large parts. To address this, we developed a data-driven spatio-temporal model based on transformer architecture to predict the location-dependent mechanical properties based on the thermal history of fabricated parts with multiple contours. The framework enables the dynamic emissivity calculation of the part for various temperatures and layer ranges to reduce the error of thermal history acquisition. We systematically compared the proposed approach's performance with other machine learning methods. The results demonstrate that the framework achieves good prediction capabilities using a small dataset. It provides a state-of-the-art methodology for predicting the spatial and temporal evolution of mechanical properties leveraging the transformer architecture. Finally, for model prediction interpretation, we investigated the location-aware morphology with various thermal profiles and mechanical properties, which allowed us to explain the reason behind each prediction.