Explainable Computer Vision Framework for Automated Pore Detection and Criticality Assessment in Additive Manufacturing

Internal porosity remains a critical defect mode in additively manufactured components, compromising structural performance and limiting industrial adoption. Automated defect detection methods exist but lack interpretability, preventing engineers from understanding the physical basis of criticality predictions. This study presents an explainable computer vision framework for pore detection and criticality assessment in three-dimensional tomographic volumes. Sequential grayscale slices were reconstructed into volumetric datasets, and intensity-based thresholding with connected component analysis identified 500 individual pores. Each pore was characterized using geometric descriptors including size, aspect ratio, extent, and spatial position relative to the specimen boundary. A pore interaction network was constructed using percentile-based Euclidean distance criteria, yielding 24,950 inter-pore connections. Machine learning models predicted pore criticality scores from extracted features, and SHAP analysis quantified individual feature contributions. Results demonstrate that normalized surface distance dominates model predictions, contributing more than an order of magnitude greater importance than all other descriptors. Pore size provides minimal influence, while geometric parameters show negligible impact. The strong inverse relationship between surface proximity and criticality reveals boundary-driven failure mechanisms. This interpretable framework enables transparent defect assessment and provides actionable insights for process optimization and quality control in additive manufacturing.

💡 Research Summary

The paper presents an explainable computer‑vision framework for automatically detecting internal pores in additive‑manufactured components and assessing their criticality. High‑resolution computed‑tomography (CT) scans are assembled into a three‑dimensional volume, and intensity‑based thresholding followed by connected‑component labeling isolates bright regions that correspond to pores. After discarding the largest component (the specimen boundary), 500 individual pores are retained. For each pore, six descriptors are extracted: (1) volume (voxel count), (2) aspect ratio, (3) extent (fraction of the bounding box occupied), (4) axial position (Z‑coordinate), (5) normalized surface distance (minimum distance from the pore centroid to the external surface, scaled by specimen dimensions), and (6) network connectivity derived from pairwise Euclidean distances. A percentile‑based distance cutoff (the lowest 20 % of all distances) is used to create a pore interaction network comprising 24 950 edges, ensuring a reproducible network density without an arbitrary global threshold.

A supervised regression model (implemented with ensemble methods such as Random Forest and Gradient Boosting) is trained on these features to predict a scalar criticality score for each pore. Standard data‑splitting and cross‑validation procedures are employed to avoid overfitting. To interpret the model, Shapley Additive exPlanations (SHAP) are calculated for every prediction. The mean absolute SHAP values reveal that normalized surface distance dominates the model, contributing more than an order of magnitude higher importance than any other descriptor. The SHAP beeswarm plot shows a clear monotonic inverse relationship: pores closer to the surface receive large positive SHAP values and higher predicted criticality, whereas deeper pores receive negative SHAP values and lower scores. Pore size exhibits a modest positive effect but its SHAP magnitude is far smaller, indicating a secondary, conditional role. Aspect ratio, extent, and axial position cluster around zero SHAP values, confirming negligible influence on the predictions.

These findings align with physical intuition: surface‑adjacent pores experience higher stress concentration, easier access to environmental factors, and boundary‑condition effects, making them more likely to initiate failure. In contrast, traditional emphasis on pore size or shape alone does not capture the dominant risk factor in the examined dataset.

The authors argue that the framework provides practical benefits for additive‑manufacturing quality assurance. Automated detection and criticality scoring reduce inspection time, while SHAP‑based explanations give engineers transparent insight into why a specific pore is deemed critical, enabling targeted process adjustments (e.g., laser power, scan strategy) and inspection prioritization near the surface. The network representation also opens avenues for studying pore‑pore interactions.

Future work is suggested to extend the methodology across different alloys, process parameters, and loading conditions, to validate the generality of the surface‑driven criticality mechanism, and to integrate the pipeline with real‑time CT acquisition or graph‑neural‑network models for even richer defect analysis.