Page 38 - IJAMD-1-2

International Journal of AI for
Materials and Design
AI-driven quality assurance in AM

4.4.2. Future work and data-sharing initiatives to facilitate the training and

• Advanced architectures: Exploring more advanced validation of AI models for AM quality assurance.
deep learning architectures, such as residual networks 5.2. Model interpretability and explainability
and generative adversarial networks, to improve defect
detection accuracy. The black-box nature of many AI algorithms, particularly
• Integration with other AI techniques: Combining deep learning models, presents challenges for model
CNNs with other AI techniques like RL for adaptive interpretability and explainability in AM quality
process control and generative models for defect assurance. 32,35 Understanding how AI models arrive at
prediction. their decisions and predictions is critical for gaining trust
• Scalability: Extending the methodology to other AM and acceptance from users, stakeholders, and regulatory
techniques and materials to create a more versatile agencies. Interpretable AI techniques, such as attention
quality assurance system. mechanisms, feature attribution methods, and model
visualization tools, can help elucidate the underlying
4.5. Conclusion of the case study factors driving model predictions and highlight areas of

This case study demonstrates the effectiveness of CNNs uncertainty or ambiguity. 35,36 Enhancing the interpretability
in enhancing quality assurance in SLM. By integrating and explainability of AI-driven quality assurance systems
real-time defect detection into the manufacturing process, will be essential for fostering transparency, accountability,
significant improvements in part quality and production and confidence in their deployment across AM workflows.
efficiency were achieved. This approach showcases the 5.3. Integration with manufacturing workflows
transformative potential of AI-driven methodologies in
AM, highlighting the need for continued research and Deploying AI-driven quality assurance systems into real-
development in this area. world manufacturing environments requires seamless
integration with existing workflows, processes, and
5. Challenges and future directions production systems. However, transitioning from research
36
As AM continues to evolve and gain traction across prototypes to practical implementations poses challenges
various industries, addressing ongoing challenges and related to interoperability, scalability, and compatibility
charting a course for future advancements in quality with diverse AM machines, software platforms, and data
assurance remains paramount. In this section, we explore formats. Standardizing interfaces, protocols, and data
32
the multifaceted challenges facing AI-driven quality exchange formats can facilitate interoperability and enable
assurance in AM processes and outline potential future the seamless integration of AI-driven quality assurance
directions for research and development. The integration tools into AM ecosystems. 10,19,23 Moreover, developing
of AI techniques into AM has the potential to revolutionize user-friendly interfaces, workflow automation tools, and
quality assurance processes. Table 4 explores the advantages decision support systems can streamline the adoption and
and limitations of various AI techniques, including ML, deployment of AI technologies in manufacturing settings.
computer vision, and neural networks, to provide valuable 5.4. Physics-informed AI models
insights for selecting appropriate methodologies for
practical applications. Integrating physics-based models and domain knowledge
into AI-driven quality assurance systems can enhance their
5.1. Data availability and quality robustness, generalization, and predictive capabilities.
One of the primary challenges in AI-driven quality Physics-informed AI models leverage insights from
assurance for AM is the availability and quality of training materials science, process engineering, and computational
data. Building accurate and robust AI models requires modeling to constrain and guide AI algorithms, enabling
large labeled datasets encompassing diverse defect types, them to capture underlying physical principles and
36
materials, process parameters, and printing conditions. 33,34 phenomena governing AM processes. By combining
However, acquiring and annotating such datasets can data-driven learning with mechanistic understanding,
be time-consuming, costly, and resource-intensive, physics-informed AI models can improve defect
particularly for rare or complex defects. Moreover, ensuring prediction, process optimization, and design optimization
the consistency and reliability of labeling annotations is in AM. 29,31,35,36 Future research directions should focus on
crucial for model generalization and performance. 18,19 developing hybrid AI-physical models that leverage the
Future research efforts should focus on developing complementary strengths of data-driven and physics-
standardized datasets, data augmentation techniques, based approaches for AM quality assurance.

Volume 1 Issue 2 (2024) 32 doi: 10.36922/ijamd.3455