Temporal convolutional networks for data-driven thermal modeling of directed energy deposition

Metal additive manufacturing (AM) involves complex multiscale and multiphysics processes. Physics-based modeling approaches to simulate such processes face challenges in their predictions due to the several time and length scales involved in the thermomechanical effects that are inherent in AM. Deep learning-based ap-proaches have been recently explored to address this issue, as they have been shown to be capable of capturing highly nonlinear relations between input and output features. This investigation proposes the use of temporal convolutional networks (TCNs) for fast inferencing of thermal histories in AM processes. TCNs have been pre-viously shown to be superior to other deep learning approaches while requiring less training time. A method-ology, therefore, of using TCNs in thermal history predictions for the case of directed energy deposition (DED) is presented herein. The results were found to be of comparable accuracy to other deep learning methods that have been proposed for similar predictions but at a fraction of their compute and training times.

Automated summary

Metal additive manufacturing (AM) involves complex multiscale and multiphysics processes. Deep learning-based approaches, specifically temporal convolutional networks (TCNs), have been proposed as a solution to the challenges faced by physics-based modeling methods in predicting thermal histories in AM. This study presents the use of TCNs for fast inferencing in directed energy deposition (DED) processes, achieving comparable accuracy to other deep learning methods with significantly reduced compute and training times.