Carter Taylor

Reliable in-situ detection of keyhole collapse in Laser Powder Bed Fusion (LPBF) remains challenging due to the transient nature of the melt pool and the inherent stochasticity of the process. This study presents a physically informed, machine learning-based framework that identifies collapse events directly from Thermal Energy Density (TED) obtained from coaxial photodiodes with a sampling rate of 200 kHz. Power Spectral Density (PSD) analysis of the TED signal revealed that low-frequency regions are dominated by pink noise and high-frequency regions by white noise, with bulk melt pool motion and collapse dynamics confined to the intermediate band. Accordingly, a 0.25-30 kHz band- pass filter was applied to the TED signal to isolate the melt pool dynamics. Statistical and frequency- domain features from the filtered TED are used to train a Random Forest anomaly detection model using ground truth data from the Advanced Photon Source, where high-speed operando X-ray imaging verified collapse events. Applied to full build TED data from a commercial DMG MORI LASERTEC 12 SLM (LPBF) machine, the framework generalizes within the same material system and sensing modality without retraining, preserving a strong correlation between predicted anomalies and part-level CT-measured porosity. By integrating physics-based filtering, feature-driven learning, and adaptive thresholding, this method provides a scalable and interpretable foundation for real-time LPBF defect detection and monitoring.