Explainable AI for non-destructive testing in additive manufacturing: concept bottleneck models linking ultrasound and eddy current signatures to interpretable defect mechanisms

Additive manufacturing (AM) of metals is increasingly used to produce high-performance components for aerospace, medical, and energy applications. However, laser-powder bed fusion (L-PBF) can generate process-induced defects such as lack of fusion, keyhole porosity, micro-cracks, and surface balling, which may reduce fatigue life and structural integrity. Ultrasound testing (UT) and eddy current testing (ECT) offer promising non-destructive testing solutions, but many machine-learning classifiers remain opaque and provide limited physical traceability for certification-critical applications. This paper presents a Multi-Modal Concept Bottleneck Model (MM-CBM) for interpretable AM defect diagnosis using fused UT and ECT signals. The proposed framework performs two-stage inference: first, a physics-aware concept layer predicts interpretable attributes, including backwall attenuation, phase deviation, time-of-flight shift, signal scattering, and defect depth; second, a linear classifier assigns the final defect class using only these concepts. Experiments on 12,000 synchronized UT-ECT signal pairs from Inconel 718 L-PBF specimens show that MM-CBM achieves 93.5% accuracy and an AUC of 0.978 while providing auditable, concept-level explanations for each diagnosis.