Temperature evolution prediction for laser directed energy deposition enabled by finite element modelling and bi-directional gated recurrent unit

Abstract

In the laser-directed energy deposition (L-DED) process, achieving an efficient temperature evolution prediction of molten pools is critical but challenging. To resolve this issue, this study presents an innovative approach that integrates a high-fidelity finite element (FE) model and an effective machine-learning model. Firstly, a high-fidelity FE model for the L-DED process was developed and subsequently validated through an experimental examination of the cross-sectional geometries of the molten pools and temperature fields of the substrate. Then, a Bi-directional gated recurrent unit (Bi-GRU) was formulated to predict the temperature evolution of the molten pools during L-DED. By training the Bi-GRU model using datasets generated from the FE model, it was deployed to efficiently predict the temperature evolution of the manufactured multi-layer single-bead walls. The results demonstrated that, in terms of the average mean absolute error, this approach outperformed other approaches designed based on the gated recurrent unit (GRU) model, long short-term memory model, and recurrent neural network models by 26.7%, 52.1%, and 65.2%, respectively. The results also showed that the prediction time required by this approach, once trained, was significantly reduced by five orders of magnitude compared with the FE model. Therefore, this approach accurately predicts the temperature evolution of multi-layer single-bead walls in a computationally efficient manner. This approach is a promising solution for supporting the real-time control of the L-DED process in industrial applications.