A novel dual-phase Bayesian estimation and multi-task learning method for process optimization for reducing lack-of-fusion defects during laser powder bed fusion

In metal additive manufacturing, lack-of-fusion is one of the most critical porosity defects. Its complex formation mechanism makes the modeling process highly challenging. While existing researches have focused on high-fidelity simulations, high-performance predictions and process map based on single-track scenarios, the actual printing process requires consideration of the interlapping and stacking of multi-layer and multi-track melt pools. This significantly increases the computational cost of high-fidelity simulations and the computational error of traditional analytical models. In this study, we developed the dual-phase Bayesian estimation and multi-task learning (Dual-BE&ML) method. This approach innovatively “teaches” the machine learning models to account for system error and uncertainty by incorporating physical laws. It also demonstrates enhanced fitting capabilities for the melt pool width. Using a set of dimensionless numbers, we constructed a sensitivity map for inter-layer lack-of-fusion porosity in laser powder bed fusion of 316 L stainless steel. This allows us to accurately avoid process regions prone to inter-layer lack-of-fusion porosity during the process optimization. The results show that when the ratio of melt pool depth to layer thickness exceeds 1.62 and the ratio of hatch spacing to melt pool width is less than 0.76, the inter-layer lack-of-fusion porosity disappears. This not only confirms the reliability of our approach but also provides important guidance for accelerating process optimization and product design.