DCGAN-CNN with physical constraints for porosity prediction in laser metal deposition with unbalanced data

Abstract

The recent explosion of physics-informed machine learning in additive manufacturing (AM) process modeling has shown the complementary strengths of data-driven models and physics-based models. Rich measured thermal images from the melt pool provide opportunities to develop unique methods for porosity analytics, but the data quality limits the performance of predictive models. To deal with unbalanced data, data augmentation is required, but such surrogate data is likely to be physically invalid, leading to inaccurate predictions. In this paper, a deep learning model with physical constraints is developed to predict porosity in laser metal deposition (LMD). The model leverages deep convolutional generative adversarial network (DCGAN) for data generation. It is also integrated with a Kullback-Leibler (KL) divergence-based model that characterizes the shape and temperature distribution of the melt pool and a control chart that ensures the augmented data adheres to the physical characteristics. The porosity label is predicted by a convolutional neural network (CNN) using augmented data. Experimental results demonstrate that the prediction performance outperforms benchmark deep learning methods.