Comparison of machine learning based methods on prediction quality of thin-walled geometries using laser-based Direct Energy Deposition

Abstract

Laser-based Direct Energy Deposition of small parts with high build-up rates deals with complex relationships of the process parameters and constantly varying boundary conditions. To avoid dimensional deviations during the build-up, the process parameters must be adjusted. These geometric deviations are based on the thermal conditions determined by the energy input of the previous layers. Machine learning algorithms can be used to identify appropriate process parameters, enabling economic and resource-efficient process development. The aim of this work is to compare LSTM (Long Short-Term Memory) and XGBoost (Extreme Gradient Boosting) as single model and model chain, regarding their prediction accuracy of the laser power for the layered wall build-up with a constant wall width. The model performance is calculated based on newly manufactured test samples. The results show that the chained LSTM has the highest deviation, whereas the XGBoost algorithm proves to be the most accurate.