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

Abstract: 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.

• The paper introduces an explainable vision framework that automates pore detection and quantifies defect criticality, highlighting normalized surface distance as the dominant factor.
• It employs 3D tomographic data, intensity-based thresholding, and connected-component analysis combined with SHAP values for transparent machine learning interpretation.
• The findings advocate for surface-focused quality control strategies in metal additive manufacturing, urging adaptive scanning and process optimization to mitigate critical surface pores.

Explainable Computer Vision for Automated Pore Criticality Assessment in Additive Manufacturing

Introduction

Internal porosity remains one of the most consequential defect modes in metal additive manufacturing (AM), directly degrading the structural performance, lifetime, and reliability of printed components. While advances in both computed tomography (CT) imaging and automated defect detection via computer vision and ML have partially addressed the burden of high-throughput analysis, the lack of physical interpretability in predictive models has hindered widespread industrial adoption. Purely data-driven methods obscure the mechanisms by which pores are classified as critical or benign, frustrating engineering efforts to optimize processes or implement risk-based quality assurance. The studied paper presents an explainable computer vision framework that not only automates pore detection and quantifies criticality but also elucidates the underlying physical factors driving model inference using SHAP (SHapley Additive exPlanations) analysis (2602.03883).

Technical Approach

Three-dimensional tomographic data are acquired as sequential grayscale slices and reconstructed into volumetric representations of AM specimens. Intensity-based thresholding is performed to isolate high-contrast regions, corresponding to the specimen boundary and potential pores, followed by connected-component analysis to segment individual pores. Pores are filtered by voxel count to exclude imaging artifacts and boundary-dominant regions, resulting in a set of 500 distinct pores for downstream analysis.

Each pore is described by a feature vector including:

• Centroid position (3D spatial location)
• Pore size (voxel count)
• Aspect ratio and extent (morphology descriptors)
• Axial position within the volume
• Normalized distance from surface

A pore interaction network is constructed, linking pores within the lowest 20th percentile of Euclidean centroid distances, ensuring consistency and reproducibility of network density across datasets.

A supervised machine learning model is trained using these descriptors to predict a scalar pore criticality score. Critically, model explainability is realized through application of SHAP values to quantify the local and global importance of each feature, thereby associating pore and network structure descriptors with model predictions in a transparent, non-model-specific fashion.

Results and Quantitative Findings

A total of 500 pores were segmented, and 24,950 inter-pore connections were established following percentile-based distance thresholding. The spatial distribution analysis revealed a pronounced shell-dominated morphology, with pores strongly localized near the specimen’s radial boundary rather than being uniformly dispersed.

Model introspection via SHAP yielded several statistically significant, and in some cases, counterintuitive findings:

• Normalized surface distance dominates pore criticality predictions, exceeding the mean absolute SHAP value of all other features by more than an order of magnitude.
• Pore size, despite its presumed relevance to mechanical integrity, exerts minimal influence on the model outcome. Large and small pores alike span the full range of criticality scores, and SHAP analysis indicates that pore size acts only as a weak, positive modifier (secondary effect).
• Morphological descriptors such as aspect ratio, extent, and axial (Z) position contribute negligibly, with SHAP values tightly clustered near zero.
• A robust, nearly linear inverse relationship exists between normalized surface distance and predicted criticality score. Pores proximate to the surface uniformly register higher criticality, in contrast to bulk or core-embedded defects.
• The learned response of the model can be effectively summarized as $f(x) \approx$ surface distance + $\epsilon$, where $\epsilon$ aggregates all secondary terms, confirming the deterministic nature of surface proximity as a criticality control parameter.

These findings implicate boundary-driven mechanisms—stress concentration, enhanced accessibility, and surface-limited boundary conditions—as the dominant factors in AM defect criticality. The results require reconsideration of traditional quality control strategies that prioritize large pores irrespective of location.

Implications and Broader Impact

The framework fundamentally advances the implementation of explainable AI for structural integrity analysis in AM, enabling direct traceability from model output to physical attributes—a requirement for risk-mitigation in aerospace, biomedical, and other high-consequence domains. The quantitative dominance of surface proximity over intrinsic morphology in criticality assessment supports the adoption of adaptive scanning and inspection regimes that emphasize near-surface defect mitigation.

Quality assurance protocols derived from this framework can prioritize process window optimization, energy input modulation, and post-processing specifically targeted to suppress surface-adjacent pores, potentially improving component durability and reliability beyond what is achievable by global porosity reduction alone. The approach also streamlines the tuning and validation of generative models of AM, informing surrogate modeling, and simulation-driven process control.

On a theoretical level, the results reinforce the need for defect-centric, spatially aware models of failure in heterogeneous materials. Incorporation of pore network topology and process-induced anisotropy, particularly as AM complexity and application scope escalate, is further motivated by these findings.

Future research directions include extending the explainable framework to multi-material AM, other critical alloy systems, or fundamentally different defect modes (e.g., cracks, inclusions); generalizing to variable process parameters and thermal histories; and correlating criticality analysis with empirical fatigue/lifetime data for model validation and process feedback.

Conclusion

The presented explainable computer vision framework for pore detection and criticality assessment in AM demonstrates that surface proximity is the overwhelmingly dominant feature in determining the structural risk associated with internal porosity, whereas pore size and morphology are of minimal predictive value. This finding prompts a paradigm shift toward surface-focused inspection and remediation in AM quality control and establishes a pathway for transparent, mechanism-based defect assessment leveraging interpretable AI. The framework’s adaptability and transparency underscore its utility in both theoretical investigations and industrial practice, offering a foundation for next-generation process optimization in additive manufacturing.