Rapid accomplishment of cost-effective and macro-defect-free LPBF-processed Ti parts based on deep data augmentation

Abstract

Machine learning methods can accurately predict the density of as-built parts by laser power bed fusion (LPBF), providing a reference for optimizing process parameters. However, obtaining massive training data via experiments is time-consuming and high-cost. Herein, a novel data augmentation method based on the Wasserstein generative adversarial network and regularization strategy (LC-WGAN) was developed to generate new training data for density prediction modelling. Four machine learning (ML) algorithms, support vector regression (SVR), multilayer perceptron (MLP), random forest (RF) and gradient boosting decision tree (GBDT), were adopted to construct the density prediction models. Results show that the proposed LC-WGAN effectively enhanced the prediction performance of all four models. R2 of MLP dramatically increased by 73.06 %, while that of RF and GBDT was marginally improved by 10.07 % and 7.48 %, respectively. RF trained by augmented dataset has the highest prediction performance on test set with MAE, RMSE and R2 of 0.7589, 0.9584 and 0.9822, respectively, which is suitable for density prediction. Additionally, SHAP analysis reveals that energy density is the most important parameter with an overall positive impact on density. The proposed method can provide valuable insights for the limited sample modelling in other fields.