Online melt pool monitoring has become a key enabler for quality assurance in laser powder bed fusion (LPBF). However, two geometrical quantities that strongly influence defect formation (melt pool depth and layer height) are difficult to measure directly during processing and are typically obtained via time-consuming destructive metallography. This study proposes a data-driven virtual sensing framework to predict melt pool depth and layer height for single-track LPBF of AlSi10Mg by fusing coaxial high-speed imaging features with process parameters. A coaxial high-speed camera (10 kHz) is integrated into the LPBF system to capture melt pool images, from which features describing melt pool size, shape, intensity, and texture are extracted and combined with laser power and scanning speed. We benchmark three regression models—support vector regression (SVR), random forest (RF), and a deep neural network (DNN)—under different input configurations and then retrain the selected model on the full dataset of 330 tracks using an 80/20 train–test split. The combined-input DNN consistently outperforms SVR and RF for both targets, achieving higher R 2 and lower RMSE, and a mode-aware optimization strategy further reduces extreme prediction errors near regime-transition boundaries (non-fusion and keyhole-like conditions). The proposed framework enables accurate, interpretable prediction of melt pool geometry from in-situ monitoring signals, supporting process-window design and reducing reliance on destructive measurements.
Zhu et al. (Thu,) studied this question.