Spatio-Temporal Transformer for Temperature Profiles Prediction in Large Format Additive Manufacturing

In Large Format Additive Manufacturing (LFAM), there are many challenges in the design phase, especially in determining a robust printing strategy, which involves multifaceted aspects such as temperature control during the printing process. Generative design has expanded the potential for creating intricate geometries and part consolidation. However, this method can only provide feasible designs and can not predict the thermal dynamics of print surface during the printing process, thus not able to offer the optimal printing strategy. Finite element analysis and physics based model have been used extensively for offline thermal profile prediction. However, they are either too expensive to built or relying on strong assumptions to provide low fidelity prediction. To bridge this gap, this paper introduces a transformer-based predictive model aiming at accurately determining temperature profiles for an arbitrary design geometry. Knowing the distribution and variation of temperature will allow for making tailored printing strategy. The in-situ thermal data sets from several test prints are utilized to train the model. The results indicate our model successfully handles spatial and temporal complexities to achieve highly accurate predictions. The model's predictive capability provides a robust tool for the design of printing strategy prior to manufacturing.