Microstructural Control in PBF-LB via Machine Learning

Sammanfattning

Powder Bed Fusion by Laser Beam (PBF-LB) is a versatile additive manufacturing process, yet optimizing its parameters to ensure high material quality often involves a costly and time-consuming trial-and-error approach. This thesis proposes a data-efficient machine learning framework to predict and optimize the porosity and hardness of 316L stainless steel components based on the process parameters laser power, scan speed, and hatch distance. The study develops and compares two Gaussian Process models in parallel, a single-task and a multi-task model, in which the multi-task model leverages the inherent correlation between porosity and hardness to improve accuracy. To overcome the challenge of limited experimental data, an active learning strategy was implemented using Bayesian optimization to select the best data points for the models. The models were further enhanced through physics-informed machine learning by implementing feature engineering in the form of melt pool dimensions derived from the analytical Rosenthal equation. Results demonstrate that the multi-task model significantly outperforms traditional single-task architectures, achieving peak predictive accuracy (R² > 0.84) for both targets while requiring only ~ 20 training samples. The integration of physics-based features proved critical in improving model performance. Finally, multi-objective optimization using a utopia-point approach identified parameter combinations that maintained high material quality (porosity < 0.1%, hardness > 230 HV1), while also allowing optimizing for build speed. This framework provides a robust and cost-efficient way of optimizing process parameters for new materials in industrial additive manufacturing.