A physics-informed machine learning approach for temperature field prediction in metallic additive manufacturing

Abstract

Advancements in Machine Learning (ML) have provided an efficient solution for elucidating the relationships of process-structure-property in the field of metallic Additive Manufacturing (AM). Yet, the reliance on data-driven ML imposes impediments in terms of model interpretability and flexibility. This study addresses these limitations by providing a physics-informed ML strategy for thermal modelling in metallic AM. The models were developed to employ a physics-informed 3D convolutional autoencoder, integrating the physical mechanisms of heat transfer in latent space, to simulate dynamic thermal patterns during cooling and deposition phases of wire arc additive manufacturing. The physics-informed ML model was trained by finite element method thermal simulations and the performance of the developed model was evaluated through an ablation study by comparing it with two additional ML strategies. One thin-wall and three cubic structures with different deposition paths were built experimentally and modelled to test the model’s performance. Results demonstrate that the proposed physics-informed ML model has shown a rapid convergence rate and high accuracy in predicting temperature fields during the metallic AM processes, with a root mean square error within 2 degrees in the cooling phase and 60 degrees in the deposition phase. The introduction of physical constraints into the latent space can significantly enhance feature recognition and flexibility, thereby fostering a more meaningful physical convergence.