AI model achieves high accuracy in identifying source of 3D-printed parts

Researchers at the University of Illinois have shown that every industrial 3-D printer leaves a subtle, machine-specific surface pattern. A convolutional network trained on those patterns can tell which printer made a part with almost perfect accuracy.

The team produced 9,192 parts on 21 commercial machines covering four additive-manufacturing processes: digital light synthesis, multi-jet fusion, stereolithography and fused-deposition modeling. Each part was scanned on a flatbed document scanner at 5.3 µm per pixel, creating a high-resolution image library for model training and testing.

Using an EfficientNet-V2 architecture and a voting scheme across multiple random image crops, the model identified the source printer for unseen parts with 98.5 percent accuracy. It also recognized the manufacturing process and material with up to 100 percent accuracy and even inferred the build-tray position of digital-light-synthesis parts to within roughly 5 cm (~1.97 in).

The study mapped how accuracy depends on image resolution and crop size. For processes such as digital light synthesis, a 200 µm-square crop was enough; fused-deposition parts needed larger regions (≈3 mm) but tolerated lower resolution, making the method compatible with off-the-shelf cameras and scanners.

Beyond basic classification, the approach offers a practical tool for supply-chain oversight. It can confirm that a contractor used the agreed machine, flag unreported process changes, and help trace defective or counterfeit parts without embedded labels or cooperation from the supplier.