Machine Learning-Based Prediction of Porosity Formation in Laser Powder Bed Fusion of Mechanical Components

Abstract

Porosity remains one of the most persistent quality barriers in laser powder bed fusion (LPBF), especially for mechanical components that must satisfy fatigue, leak-tightness, and structural reliability requirements. This paper presents a machine learning framework for predicting LPBF porosity formation using process parameters, powder descriptors, and physically meaningful features such as volumetric, linear, and areal energy density. The study uses a literature-constrained benchmark dataset to demonstrate the workflow when proprietary in-situ and X-ray computed tomography data are not yet available. The modeling strategy compares logistic regression, support vector machine, gradient boosting, random forest, and neural-network classifiers for high-porosity risk prediction, while a random-forest regressor estimates porosity percentage. The results indicate that tree-based ensemble learning provides the strongest balance between accuracy, recall, and interpretability, with volumetric energy density, scan speed, oxygen level, and powder morphology emerging as influential predictors. The discussion connects these model behaviors to lack-of-fusion, keyhole, spatter, and gas-entrapment mechanisms. Overall, the manuscript argues that reliable porosity prediction should not rely on machine settings alone. A more useful approach combines physics-guided feature engineering, sensor-informed data streams, cross-validated learning, and explainability tools to support inspection prioritization and eventual closed-loop process control.