| Abstract Scope |
We benchmark AI/ML methods for real-time melt pool monitoring in laser powder bed fusion (LPBF), where inference latency and limited labeled data constrain deployable model design. Using 1,200 balanced melt pool images of Nickel superalloy 625 from the NIST AMMT platform, we frame anomaly detection as binary image classification and compare five models: three transfer learning architectures (ResNet50, EfficientNetB0, MobileNetV2), a Random Forest on EfficientNetB0 feature embeddings (hybrid), and a Random Forest on raw pixels (baseline). Each model is evaluated on accuracy, precision, recall, F1, AUC, training time, inference latency, and CPU and GPU usage relevant to open-architecture LPBF machines. The hybrid EfficientNetB0 plus Random Forest achieves F1 0.9451, accuracy 0.9458, AUC 0.9904, and 1.15 ms per-image inference, outperforming pure deep learning baselines on both accuracy and latency. Pairing pretrained convolutional features with classical ensembles offers a robust, deployable route to real-time melt pool anomaly detection under data-limited LPBF conditions. |