Predicting Temperature Field for Metal Additive Manufacturing using PINN

Machine-learning-based methods are gaining traction as an alternative to numerical methods in many engineering applications. Physics-informed neural network (PINN), a self-supervised method, is particularly attractive with its unique capability of guiding the training with physical laws written in the forms of partial differential equations. Thermomechanical simulation for additive manufacturing (AM), a multi-scale, multi-physics problem could potentially benefit from the use of PINN, as demonstrated in some successful attempts in the literature. In this work, PINN is applied to different metal AM processes and several challenges that limit the robustness of PINN are observed. This paper aims to provide a summary of the observations and a preliminary attempt to account for such observations in order to pave the path for future work that aims to unleash the full promise of PINN in AM-related applications.