Multimodal deep learning for enhanced temperature prediction with uncertainty quantification in directed energy deposition (DED) process

Abstract

The accurate and reliable prediction of temperature history is crucial in meeting the ever-increasing demands for part quality and process reliability in metal additive manufacturing (AM). While many recent studies based on deep learning approaches have shown promise, they are subject to major limitations: inadequate handling of deposition strategy and insufficient consideration of uncertainty, both of which impact the prediction model performance. This work proposes a novel multimodal deep learning approach for temperature prediction with uncertainty quantification in directed energy deposition (DED) process. The proposed methodology implements multimodal data fusion, combining reproduced grayscale images of deposition strategy with numerical process variables, including process parameters, geometrical features, and printing process status. Furthermore, a novel approach for direct uncertainty estimation is introduced, inspired by object detection methods in computer vision. Through extensive comparative analyses, the proposed method outperforms conventional deterministic and probabilistic deep learning approaches, as well as state-of-the-art methods, in both temperature prediction and confidence interval estimation. An ablation study further validates the effectiveness of the proposed architecture, and quality inference based on predicted temperature distributions proves the feasibility and applicability of the proposed method for advancing process and quality control in metal AM processes.