MOSAIC algorithm analyzes keyhole dynamics 2.7x faster than YOLO for 3D printing
New segmentation model runs 19.9ms per frame on CPU, beating SAM and YOLO in precision.
In laser powder bed fusion (L-PBF), unstable keyholes—gas cavities formed by the laser—cause porosity and spatter that degrade metal 3D printed parts. High-speed x-ray imaging captures these dynamics, but existing segmentation models like SAM and YOLO are too slow or inaccurate for real-time beamline experiments. Researchers (Mathesen et al.) introduce MOSAIC, a Mobile Object Segmentation algorithm tailored for adverse imaging conditions. It delivers an average F1 score of 0.894 and precision of 0.953 across 12 unique validation samples, matching or exceeding both SAM and YOLO without requiring manual labeling or model retraining.
MOSAIC’s efficiency is its standout feature: it processes cropped frames (150x250 pixels) in just 19.9 milliseconds on a CPU, compared to 54 ms for YOLO and 5,284 ms for SAM. This 2.7x speedup over YOLO and 265x over SAM makes it viable for real-time process monitoring during active experiments. By rapidly characterizing keyhole behavior, MOSAIC could help manufacturers adjust parameters on the fly to minimize defects. The paper, submitted to ISFA 2026, includes an open-source code library, paving the way for broader adoption in additive manufacturing quality control.
- MOSAIC achieves F1 score 0.894 and precision 0.953 on 12 validation samples, outperforming SAM and YOLO.
- Runs 19.9 ms per frame on CPU—2.7x faster than YOLO and 265x faster than SAM.
- Enables real-time keyhole monitoring to reduce porosity and spatter in laser powder bed fusion 3D printing.
Why It Matters
Real-time keyhole analysis could slash porosity defects in metal 3D printing, improving part reliability and production speed.