MACHINE LEARNING MODELING FOR REAL-TIME MELT POOL MONITORING IN LASER POWDER BED FUSION ADDITIVE MANUFACTURING: A HYBRID APPROACH

In laser powder bed fusion (LPBF) additive manufacturing, real-time quality monitoring is essential to prevent defects like keyholing or lack of fusion. While deep learning models can accurately classify melt pool images, they often suffer from high computational costs and slow inference times that are incompatible with the millisecond-level requirements of industrial process control. This study investigates whether a hybrid machine learning architecture can bridge the gap between high predictive accuracy and the computational efficiency required for real-time factory-floor deployment.

The researchers developed a binary classification framework to distinguish between normal and abnormal melt pool images. They benchmarked three transfer learning architectures (ResNet50, EfficientNetB0, and MobileNetV2) against two Random Forest approaches: one trained on raw pixel data and a hybrid model that uses EfficientNetB0 to extract features before classification. The models were evaluated on a balanced dataset of 1,200 images collected from a Nickel superalloy 625 build on the NIST AMMT platform. The evaluation focused on both classification performance (F1 score, AUC) and deployment metrics, including training time, inference latency, and hardware resource usage.

The hybrid EfficientNetB0-plus-Random Forest model outperformed both purely deep learning architectures and the raw pixel baseline. It achieved an F1 score of 0.9451 and an accuracy of 0.9458, while maintaining an inference time of 1.15 ms per image. This inference speed is significantly faster than the pure deep learning models—74 times faster than EfficientNetB0 and 156 times faster than ResNet50. The results suggest that decoupling feature extraction from the classification task allows for a more efficient, robust pipeline that meets the strict latency constraints of real-time manufacturing.