Defect Classification for Additive Manufacturing with Machine Learning.

Area of Science:

• Materials Science
• Mechanical Engineering
• Computer Science

Background:

• Additive manufacturing (AM) provides design flexibility but can introduce internal defects like pores.
• These defects, including keyhole and lack of fusion pores, significantly impact component mechanical properties.
• Defect type, size, and morphology influence mechanical performance, with components often exhibiting multiple defect types.

Purpose of the Study:

• To develop a robust classification model for identifying and differentiating internal defects in metal AM components.
• To assess the performance of a random forest model against unsupervised methods for defect classification.
• To provide a practical tool for defect analysis without requiring in situ monitoring.

Main Methods:

• A random forest tree model was developed to classify defects from binary images of micrographs.
• The model was trained to distinguish between keyhole defects, lack of fusion defects, and process pores.
• Performance was evaluated by comparing classification accuracy with unsupervised learning models.

Main Results:

• The random forest model achieved approximately 95% accuracy in classifying keyhole, lack of fusion, and process pores.
• Unsupervised models demonstrated significantly lower prediction accuracies, below 60%.
• Classification accuracy for differentiating lack of fusion and keyhole defects was influenced by manufacturing parameters due to keyhole defect irregularity.

Conclusions:

• The developed random forest model offers a highly accurate method for classifying critical internal defects in additive manufacturing.
• This approach provides a valuable tool for quality assessment and process optimization in metal 3D printing.
• The model's effectiveness highlights the potential of machine learning in analyzing microstructural defects without advanced monitoring systems.