Explainable AI frameworks using SHAP and LIME enhance interpretable defect classification in additive manufacturing

Although machine learning models can detect additive manufacturing (AM) defects with over 90% accuracy, their black-box nature limits adoption in aerospace, medical, and defence sectors, where transparency, traceability, and regulatory compliance are essential. This study presents a hybrid explainable AI framework that systematically combines SHAP and LIME for interpretable AM defect classification. Using 8,420 samples spanning four defect categories from three AM processes—laser powder bed fusion, fused deposition modelling, and directed energy deposition—the framework integrates SHAP’s stable global explanations for compliance documentation with LIME’s rapid local explanations for real-time quality control. The proposed method achieves 94.3% classification accuracy, outperforming the baseline ensemble of Random Forest, VGG-16, ResNet-50, and a standard non-XAI ensemble by 6.4 percentage points. It also reduces explanation time by 85%, from 2914.5 ms with SHAP alone to 434.1 ms, corresponding to a 6.71× speedup. User studies involving 25 manufacturing engineers and quality operators showed significant improvements in trust and usability (p < 0.01). The framework supports alignment with AS9100, ISO 13485, and FDA expectations, enabling credible and efficient AI deployment in safety-critical AM environments.