Trend AnalysisEngineering
Metal Additive Manufacturing: AI-Driven Defect Detection for Aerospace Quality
Metal 3D printing (additive manufacturing, AM) can produce complex titanium, nickel superalloy, and stainless steel components impossible to make by traditional machining โ internal cooling channels, ...
By Sean K.S. Shin
This blog summarizes research trends based on published paper abstracts. Specific numbers or findings may contain inaccuracies. For scholarly rigor, always consult the original papers cited in each post.
The Question
Metal 3D printing (additive manufacturing, AM) can produce complex titanium, nickel superalloy, and stainless steel components impossible to make by traditional machining โ internal cooling channels, topology-optimised lattices, consolidated assemblies. Aerospace and medical industries are adopting metal AM for production parts, but certification requires defect rates below 1 per million. Current metal AM processes (laser powder bed fusion, directed energy deposition) produce defects โ porosity, lack-of-fusion, cracking โ at rates that demand extensive post-build inspection. Can AI-driven in-situ monitoring detect and prevent defects in real-time, eliminating the need for costly post-build CT scanning?
Landscape
Imran et al. (2024) comprehensively reviewed directed energy deposition (DED) techniques, defect types, and quality monitoring approaches. They catalogued the defect taxonomy: gas porosity (spherical voids from trapped gas), lack-of-fusion (irregular voids from insufficient melting), solidification cracking (from residual stress), and delamination (from thermal mismatch between layers). Each defect type has characteristic morphology detectable by different sensing modalities.
Yin et al. (2025) applied convolutional neural networks (CNNs) to metal AM defect detection from layer-wise powder bed images, achieving >99% accuracy in automated classification of powder bed morphology defects. Their AI models demonstrate the potential for reducing reliance on human operators for quality monitoring.
J. Wang et al. (2025) reviewed machine learning approaches for image-based process monitoring across metal AM platforms, identifying that transfer learning (reusing models trained on one material/machine for another) can reduce the data requirements for deploying AI monitoring on new systems โ critical because each machine and material combination produces different baseline imagery.
Key Claims & Evidence
<
| Claim | Evidence | Verdict |
|---|
| Metal AM defects are classifiable by powder bed imaging | CNN classification of morphology defects from layer-wise images with >99% accuracy (Yin et al. 2025) | Demonstrated; accuracy depends on imaging resolution and coverage |
| AI monitoring can reduce reliance on human operators | Automated defect classification from powder bed images (Yin et al. 2025) | Supported for layer-wise monitoring; real-time melt-pool-level monitoring more challenging |
| Transfer learning reduces training data requirements | Models transfer between materials/machines with fine-tuning (J. Wang et al. 2025) | Promising; cross-machine transfer effectiveness varies |
| DED and PBF require different monitoring strategies | Different defect types and process dynamics (Imran et al. 2024) | Confirmed; platform-specific solutions needed |
Open Questions
Closed-loop control: Can AI monitoring not just detect defects but trigger real-time process parameter correction (laser power, scan speed, layer height) to prevent them?
Certification: Will aviation authorities (FAA, EASA) accept AI-monitored AM parts with reduced post-build inspection, or will CT scanning remain mandatory?
Multi-physics simulation: Can digital twin models predict defect formation from process parameters alone, enabling defect prevention before monitoring is needed?
Data standardisation: Each AM machine manufacturer uses proprietary sensor configurations and data formats. Can open standards enable cross-platform AI model development?Referenced Papers
- [1] Imran, M.M. et al. (2024). Advancements in 3D Printing: DED, Defect Analysis, and Quality Monitoring. Technologies, 12(6), 86. DOI: 10.3390/technologies12060086
- [2] Yin, X. et al. (2025). AI-driven defect detection in metal 3D printing using CNNs. Materials Science in Additive Manufacturing. DOI: 10.36922/msam025150022
- [3] Wang, J. et al. (2025). Machine learning in image-based metal AM process monitoring. eScience in Additive Manufacturing. DOI: 10.36922/esam.8548
๋ฉด์ฑ
์กฐํญ: ์ด ๊ฒ์๋ฌผ์ ์ ๋ณด ์ ๊ณต ๋ชฉ์ ์ ์ฐ๊ตฌ ๋ํฅ ๊ฐ์์ด๋ค. ํ์ ์ฐ๊ตฌ์์ ์ธ์ฉํ๊ธฐ ์ ์ ๊ตฌ์ฒด์ ์ธ ์ฐ๊ตฌ ๊ฒฐ๊ณผ, ํต๊ณ ๋ฐ ์ฃผ์ฅ์ ์๋ฌธ ๋
ผ๋ฌธ๊ณผ ๋์กฐํ์ฌ ๊ฒ์ฆํด์ผ ํ๋ค.
๊ธ์ ์ ์ธต ์ ์กฐ: ํญ๊ณต์ฐ์ฃผ ํ์ง์ ์ํ AI ๊ธฐ๋ฐ ๊ฒฐํจ ํ์ง
๋ถ์ผ: ๊ณตํ | ๋ฐฉ๋ฒ๋ก : ์คํ-๊ณ์ฐ
์ ์: Sean K.S. Shin | ๋ ์ง: 2026-03-17
์ฐ๊ตฌ ์ง๋ฌธ
๊ธ์ 3D ํ๋ฆฐํ
(์ ์ธต ์ ์กฐ, AM)์ ๊ธฐ์กด ์ ์ญ ๊ฐ๊ณต์ผ๋ก๋ ์ ์์ด ๋ถ๊ฐ๋ฅํ ๋ณต์กํ ํฐํ๋, ๋์ผ ์ดํฉ๊ธ, ์คํ
์ธ๋ฆฌ์ค๊ฐ ๋ถํ์ ์ ์กฐํ ์ ์๋ค โ ๋ด๋ถ ๋๊ฐ ์ฑ๋, ์์ ์ต์ ํ ๊ฒฉ์ ๊ตฌ์กฐ, ํตํฉ ์กฐ๋ฆฝ์ฒด ๋ฑ์ด ๊ทธ ์์ด๋ค. ํญ๊ณต์ฐ์ฃผ ๋ฐ ์๋ฃ ์ฐ์
์ ์์ฐ ๋ถํ์ ๊ธ์ AM์ ๋์
ํ๊ณ ์์ผ๋, ์ธ์ฆ์ ์ํด์๋ 100๋ง ๊ฐ๋น 1๊ฐ ๋ฏธ๋ง์ ๊ฒฐํจ๋ฅ ์ด ์๊ตฌ๋๋ค. ํํ ๊ธ์ AM ๊ณต์ (๋ ์ด์ ๋ถ๋ง ๋ฒ ๋ ์ตํฉ, ์ง์ ์๋์ง ์ ์ธต)์ ๊ธฐ๊ณต, ๋ฏธ์ฉ์ต, ๊ท ์ด ๋ฑ์ ๊ฒฐํจ์ ๋ฐ์์ํค๋ฉฐ, ์ด๋ ๊ด๋ฒ์ํ ํ์ฒ๋ฆฌ ๊ฒ์ฌ๋ฅผ ํ์๋ก ํ๋ค. AI ๊ธฐ๋ฐ ์ธ์ํ(in-situ) ๋ชจ๋ํฐ๋ง์ด ์ค์๊ฐ์ผ๋ก ๊ฒฐํจ์ ํ์งํ๊ณ ์๋ฐฉํจ์ผ๋ก์จ ๋น์ฉ์ด ๋ง์ด ๋๋ ํ์ฒ๋ฆฌ CT ์ค์บ์ ํ์์ฑ์ ์์จ ์ ์์๊น?
์ฐ๊ตฌ ๋ํฅ
Imran et al. (2024)์ ์ง์ ์๋์ง ์ ์ธต(DED) ๊ธฐ๋ฒ, ๊ฒฐํจ ์ ํ, ํ์ง ๋ชจ๋ํฐ๋ง ์ ๊ทผ๋ฒ์ ํฌ๊ด์ ์ผ๋ก ๊ฒํ ํ์๋ค. ์ด๋ค์ ๊ฒฐํจ ๋ถ๋ฅ ์ฒด๊ณ๋ฅผ ๋ค์๊ณผ ๊ฐ์ด ์ ๋ฆฌํ์๋ค: ๊ฐ์ค ๊ธฐ๊ณต(ํฌํ๋ ๊ฐ์ค๋ก ์ธํ ๊ตฌํ ๊ณต๊ทน), ๋ฏธ์ฉ์ต(๋ถ์ถฉ๋ถํ ์ฉ์ต์ผ๋ก ์ธํ ๋ถ๊ท์น ๊ณต๊ทน), ์๊ณ ๊ท ์ด(์๋ฅ ์๋ ฅ์ผ๋ก ์ธํ ๊ฒฐํจ), ์ธต๊ฐ ๋ฐ๋ฆฌ(์ธต ๊ฐ ์ด์ ๋ถ์ผ์น๋ก ์ธํ ๊ฒฐํจ). ๊ฐ ๊ฒฐํจ ์ ํ์ ์๋ก ๋ค๋ฅธ ๊ฐ์ง ๋ฐฉ์์ผ๋ก ํ์ง ๊ฐ๋ฅํ ํน์ง์ ์ธ ํํ๋ฅผ ์ง๋๋ค.
Yin et al. (2025)์ ํฉ์ฑ๊ณฑ ์ ๊ฒฝ๋ง(CNN)์ ์ธต๋ณ ๋ถ๋ง ๋ฒ ๋ ์ด๋ฏธ์ง๋ก๋ถํฐ์ ๊ธ์ AM ๊ฒฐํจ ํ์ง์ ์ ์ฉํ์ฌ ๋ถ๋ง ๋ฒ ๋ ํํ ๊ฒฐํจ์ ์๋ ๋ถ๋ฅ์์ 99% ์ด์์ ์ ํ๋๋ฅผ ๋ฌ์ฑํ์๋ค. ์ด๋ค์ AI ๋ชจ๋ธ์ ํ์ง ๋ชจ๋ํฐ๋ง์์ ์ธ๊ฐ ์ด์์์ ๋ํ ์์กด๋๋ฅผ ์ค์ผ ์ ์๋ ๊ฐ๋ฅ์ฑ์ ๋ณด์ฌ์ค๋ค.
J. Wang et al. (2025)์ ๊ธ์ AM ํ๋ซํผ ์ ๋ฐ์ ๊ฑธ์น ์ด๋ฏธ์ง ๊ธฐ๋ฐ ๊ณต์ ๋ชจ๋ํฐ๋ง์ ์ํ ๋จธ์ ๋ฌ๋ ์ ๊ทผ๋ฒ์ ๊ฒํ ํ์์ผ๋ฉฐ, ์ ์ด ํ์ต(ํ ์ฌ๋ฃ/์ฅ๋น์์ ํ๋ จ๋ ๋ชจ๋ธ์ ๋ค๋ฅธ ํ๊ฒฝ์ ์ฌํ์ฉํ๋ ๋ฐฉ๋ฒ)์ด ์๋ก์ด ์์คํ
์ AI ๋ชจ๋ํฐ๋ง์ ๋ฐฐ์นํ๋ ๋ฐ ํ์ํ ๋ฐ์ดํฐ ์๊ตฌ๋์ ์ค์ผ ์ ์์์ ํ์ธํ์๋ค โ ์ด๋ ๊ฐ ์ฅ๋น์ ์ฌ๋ฃ ์กฐํฉ์ด ์๋ก ๋ค๋ฅธ ๊ธฐ์ค ์ด๋ฏธ์ง๋ฅผ ์์ฑํ๊ธฐ ๋๋ฌธ์ ๋งค์ฐ ์ค์ํ ์ฌํญ์ด๋ค.
ํต์ฌ ์ฃผ์ฅ ๋ฐ ๊ทผ๊ฑฐ
<
| ์ฃผ์ฅ | ๊ทผ๊ฑฐ | ํ์ |
|---|
| ๊ธ์ AM ๊ฒฐํจ์ ๋ถ๋ง ๋ฒ ๋ ์ด๋ฏธ์ง์ผ๋ก ๋ถ๋ฅ ๊ฐ๋ฅํ๋ค | ์ธต๋ณ ์ด๋ฏธ์ง๋ก๋ถํฐ 99% ์ด์์ ์ ํ๋๋ก ํํ ๊ฒฐํจ์ CNN ๋ถ๋ฅ (Yin et al. 2025) | ์
์ฆ๋จ; ์ ํ๋๋ ์ด๋ฏธ์ง ํด์๋ ๋ฐ ์ปค๋ฒ๋ฆฌ์ง์ ๋ฐ๋ผ ๋ฌ๋ผ์ง |
| AI ๋ชจ๋ํฐ๋ง์ ์ธ๊ฐ ์ด์์์ ๋ํ ์์กด๋๋ฅผ ์ค์ผ ์ ์๋ค | ๋ถ๋ง ๋ฒ ๋ ์ด๋ฏธ์ง๋ก๋ถํฐ์ ์๋ ๊ฒฐํจ ๋ถ๋ฅ (Yin et al. 2025) | ์ธต๋ณ ๋ชจ๋ํฐ๋ง์ ๋ํด์๋ ์ง์ง๋จ; ์ค์๊ฐ ์ฉ์ตํ ์์ค ๋ชจ๋ํฐ๋ง์ ๋์ฑ ์ด๋ ค์ |
| ์ ์ด ํ์ต์ ํ๋ จ ๋ฐ์ดํฐ ์๊ตฌ๋์ ์ค์ธ๋ค | ๋ฏธ์ธ ์กฐ์ ์ ํตํด ์ฌ๋ฃ/์ฅ๋น ๊ฐ ๋ชจ๋ธ ์ ์ด ๊ฐ๋ฅ (J. Wang et al. 2025) | ์ ๋งํจ; ์ฅ๋น ๊ฐ ์ ์ด ํจ๊ณผ๋ ๋ค์ํจ |
| DED์ PBF๋ ์๋ก ๋ค๋ฅธ ๋ชจ๋ํฐ๋ง ์ ๋ต์ด ํ์ํ๋ค | ๊ฒฐํจ ์ ํ ๋ฐ ๊ณต์ ์ญํ์ ์ฐจ์ด (Imran et al. 2024) | ํ์ธ๋จ; ํ๋ซํผ๋ณ ์๋ฃจ์
ํ์ |
๋ฏธํด๊ฒฐ ๊ณผ์
ํ๋ฃจํ ์ ์ด: AI ๋ชจ๋ํฐ๋ง์ด ๊ฒฐํจ์ ํ์งํ๋ ๊ฒ์ ๊ทธ์น์ง ์๊ณ , ๊ฒฐํจ์ ์๋ฐฉํ๊ธฐ ์ํด ์ค์๊ฐ ๊ณต์ ๋ณ์ ๋ณด์ (๋ ์ด์ ์ถ๋ ฅ, ์ค์บ ์๋, ์ธต ๋์ด)์ ์ ๋ฐํ ์ ์๋๊ฐ?
์ธ์ฆ: ํญ๊ณต ๋น๊ตญ(FAA, EASA)์ด ํ์ฒ๋ฆฌ ๊ฒ์ฌ๋ฅผ ์ค์ธ AI ๋ชจ๋ํฐ๋ง AM ๋ถํ์ ์น์ธํ ๊ฒ์ธ๊ฐ, ์๋๋ฉด CT ์ค์บ์ด ์ฌ์ ํ ์๋ฌด์ ์ผ๋ก ์ ์ง๋ ๊ฒ์ธ๊ฐ?
๋ค์ค ๋ฌผ๋ฆฌ ์๋ฎฌ๋ ์ด์
: ๋์งํธ ํธ์ ๋ชจ๋ธ์ด ๊ณต์ ๋ณ์๋ง์ผ๋ก ๊ฒฐํจ ํ์ฑ์ ์์ธกํ์ฌ, ๋ชจ๋ํฐ๋ง์ด ํ์ํ๊ธฐ ์ด์ ์ ๊ฒฐํจ ์๋ฐฉ์ ๊ฐ๋ฅํ๊ฒ ํ ์ ์๋๊ฐ?
๋ฐ์ดํฐ ํ์คํ: ๊ฐ AM ์ฅ๋น ์ ์กฐ์ฌ๋ ๋
์์ ์ธ ์ผ์ ๊ตฌ์ฑ๊ณผ ๋ฐ์ดํฐ ํ์์ ์ฌ์ฉํ๋ค. ๊ฐ๋ฐฉํ ํ์ค์ด ํ๋ซํผ ๊ฐ AI ๋ชจ๋ธ ๊ฐ๋ฐ์ ๊ฐ๋ฅํ๊ฒ ํ ์ ์๋๊ฐ?References (3)
Imran, M. M., Che Idris, A., De Silva, L. C., Kim, Y., & Abas, P. E. (2024). Advancements in 3D Printing: Directed Energy Deposition Techniques, Defect Analysis, and Quality Monitoring. Technologies, 12(6), 86.
Yin, X., Akmal, J., & Salmi, M. (2025). Artificial intelligence-driven defect detection and localization in metal 3D printing using convolutional neural networks. Materials Science in Additive Manufacturing, 4(3), 025150022.
Wang, J., Zhang, X., & Lu, Y. (2025). Machine learning in image-based metal additive manufacturing process monitoring and control: A review. Engineering Science in Additive Manufacturing, 1(1), 8548.