Prediction of Inconel718 deposition layer geometry by fusing experimental, simulation data, and machine learning model

Laser directed energy deposition (DED) technology is used in a variety of industry sectors. It is well known that manufacturing process parameters affect deposition layer shape, which can affect specimen performance. Although machine learning has provided assistance in predicting the morphology of the deposition layer, depending exclusively on experimental results to predict the deposition layer morphology needs large resources. In this study, the simulation model with an error of less than 5% is established, and then the deposition layer characteristics are predicted using experimental data and the experimental and simulation fusion data set to reduce experimental cost and enhance prediction accuracy. An improved RIME method based on a chaotic mapping mechanism and Levy flight opposition-based learning (CM-LOBL-RIME) is proposed to improve the LSBoost algorithm. A comparative analysis is conducted with eight other models.The SHAP method is employed to evaluate the primary contributions of laser power, powder feed rate, and scanning speed to the predictive model. The results show that the improved LSBoost algorithm predicts deposition depth, height, and width with 0.90, 0.91, and 0.93 for the experimental data set and 0.92, 0.95, and 0.96 for the fusion data set. The experimental data-based prediction model had an accuracy R² of 0.923, 0.9279, and 0.942 in the test set, whereas the fusion data-based model had 0.9444, 0.9659, and 0.9715. The new dataset shows deposition layer depth, height, and width errors below 7%, 5%, and 2%, respectively. Thus, the improved machine learning models exhibited remarkable prediction accuracy, excellent generalization, and robustness. The depth and width predicted models are influenced by laser power, while the height predicted models are affected by powder feed rate.