Process-agnostic multi-metric nondestructive characterization of porosity in additive manufacturing via specialized single-camera two-wavelength pyrometry

Predicting and quantifying pores in laser powder bed fusion (LPBF) remains limited by models tailored to narrow process regimes and by conventional thermal sensing that is highly sensitive to emissivity, both of which fail under realistic process variability. Here, we introduce a process-agnostic, context-aware deep learning framework for quantitative, multi-metric pore characterization, enabled by our patented single-camera two-wavelength imaging pyrometry (STWIP) system. This novel high-speed (up to 100,000 frames per second (fps), with 22,500 fps demonstrated in the present study), high-resolution diagnostic uniquely captures spatiotemporal melt pool thermodynamics and morphology with physical fidelity, overcoming emissivity and alignment challenges inherent to conventional monitoring. By integrating frequency-domain and temporal representations across adjacent layers and laser scans, the optimized model achieves 88% accuracy, 85% precision, and 84% recall for pore-presence detection, with 83% mean accuracy for pore count and size and 72% for porosity ratio across unseen test sets. Performance remains robust, albeit slightly lower, on entirely new builds. These results are remarkably strong given the numerically small ground-truth pore metrics and the extreme data imbalance and process variability of realistic LPBF, and they substantially exceed prior qualitative or regime-specific approaches. Our method enables robust, quantitative, and potentially transferable defect prediction across regimes, advancing real-time quality assurance in metal additive manufacturing.