Page 55 - MSAM-4-3

Materials Science in Additive Manufacturing AI-driven defect detection in metal AM

false predictions. Transfer learning proved effective in indirectly. Other authors declare they have no competing
accelerating convergence and boosting performance interests.
under limited data diversity and computational
resources, and appropriate image pre-processing Author contributions
improved the detection of small-scale defects. Conceptualization: Mika Salmi and Xinyi Yin
• Given the limited exploration of object detection Data curation: Xinyi Yin
models in AM, this study investigates their potential Formal analysis: Xinyi Yin
and demonstrates that both Faster R-CNN and Methodology: Xinyi Yin and Jan Akmal
YOLOv5 can effectively localize defect regions to Software: Xinyi Yin
support human inspection. YOLOv5 displays greater Supervision: Mika Salmi and Jan Akmal
robustness to scale variation and complex shapes, Visualization: Xinyi Yin
significantly outperforming Faster R-CNN. Writing – original draft: Xinyi Yin
• To address data imbalance and the scarcity of high- Writing – review and editing: Mika Salmi and Jan Akmal
quality AM datasets, this study contributes an
annotated image dataset. The high similarity among Ethics approval and consent to participate
layer-wise images often limits model generalization Not applicable.
across defect types. Public release of the dataset aims
to increase data diversity, improve adaptability to novel Consent for publication
defects, and support the development of intelligent
AM quality monitoring. Not applicable.
• The limited precision of present object detection Availability of data
models is mainly due to the abstract nature of defects,
scale variation, low contrast with backgrounds, noise The dataset used in this study has been deposited in Zenodo
interference, and overlapping bounding boxes. Future and is publicly available at: https://doi.org/10.5281/
improvements may include dataset expansion, refined zenodo.14996806.
annotations (e.g., mask labeling), and adoption of
advanced detection frameworks to enhance accuracy References
and generalizability for industrial applications. 1. Herzog T, Brandt M, Trinchi A, Sola A, Molotnikov A.
Process monitoring and machine learning for defect
Overall, this work provides a comprehensive, task- detection in laser-based metal additive manufacturing.
aligned evaluation of CNN models for AM defect J Intell Manuf. 2024;35(4):1407-1437.
monitoring, supported by a realistic dataset and performance
benchmarks. The findings serve as a valuable reference for doi: 10.1007/s10845-023-02119-y
future research on model selection, deployment strategies, 2. Kim H, Lin Y, Tseng TLB. A review on quality control in
and data standardization in AM quality control. additive manufacturing. Rapid Prototyp J. 2018;24(3):645-669.
doi: 10.1108/RPJ-03-2017-0048
Acknowledgments
3. Vasques CMA, Cavadas AMS, Abrantes JCC.
The authors would like to thank Björkstrand Roy, the Technology overview and investigation of the quality of a
laboratory manager of ADDLAB at Aalto University, for 3D-printed maraging steel demonstration part. MSAM.
providing the raw data used in this study. 2025;4(2):025040002.

Funding doi: 10.36922/msam025040002
4. DebRoy T, Wei HL, Zuback JS, et al. Additive manufacturing
This research was financially supported by the Finnish of metallic components-process, structure and properties.
Doctoral Program Network in Artificial Intelligence Prog Mater Sci. 2018;92:112-224.
(AI-DOC, decision number VN/3137/2024-OKM-6) and doi: 10.1016/j.pmatsci.2017.10.001
the Tandem Industry Academia funding from the Finnish
Research Impact Foundation. 5. Tan C, Li R, Su J, et al. Review on field assisted metal additive
manufacturing. Int J Mach Tools and Manuf. 2023;189:104032.
Conflict of interest doi: 10.1016/j.ijmachtools.2023.104032
Mika Salmi serves as the Editorial Board Member of the 6. Brennan MC, Keist JS, Palmer TA. Defects in metal
journal but was not in any way involved in the editorial and additive manufacturing processes. J Mater Eng Perform.
peer-review process conducted for this paper, directly or 2021;30(7):4808-4818.

Volume 4 Issue 3 (2025) 13 doi: 10.36922/MSAM025150022