Predicting porosity in laser powder bed fusion of metals (PBF-LB/M) using scanning data and machine learning

• Embedded high-speed scan-data acquisition in a commercial PBF-LB/M system to derive hotspot features and classify high-risk Spatter/Lack-of-Fusion and Keyhole regions. • Machine learning models trained on scan-derived hotspots and XCT ground truth, with LSBoost achieving MAE ≈ 7.29, RMSE ≈ 11.68, R² ≈ 0.76, r ≈ 0.88 for porosity prediction. • Demonstrated that scan data coupled with ML enables in situ identification of porosity-prone layers in PBF-LB/M, with improved prediction for pores larger than 0.100 mm. Laser powder bed fusion of metals (PBF-LB/M) remains prone to process-induced porosity, and practical tools that connect scan behavior to defect formation are still limited. In this study, we present a framework that predicts porosity from scan data acquired on a commercial PBF-LB/M system using an embedded high-speed scan acquisition device. The device records scanning position, time, and laser power at high temporal resolution, enabling reconstruction of layer-wise exposure paths that are normally inaccessible to users. From these data, we derive porosity-prone “hotspot” regions based on high linear energy input, energy input gradients, and deceleration or acceleration events along the scan path. Ground-truth porosity is obtained from X-ray Computed Tomography (XCT) of a compact qualification artifact containing multiple geometries, including lattice structures. Across the machine learning evaluated models, LSBoost provides the best overall performance when predicting total pore counts per layer with a Mean Absolute Error (MAE) ≈ 7.29. Prediction of larger pores with equivalent diameters greater than 0.100 mm was slightly more accurate (MAE = 4.747), indicating stronger correlation for critical part anomalies. However, porosity outliers tend to be underestimated, highlighting both the need for improved calibration and the benefit of additional in situ process signals that capture interactions with other process variables such as powder or gas flow. Overall, the results demonstrate that scan-derived hotspot features, combined with machine learning, are a viable basis for in situ identification and prediction of porosity-prone layers in PBF-LB/M.