<![CDATA[Marine welding defects threaten structural reliability because conventional inspection and black-box prediction models can identify discontinuities without adequately explaining their physical origins. This study aimed to develop an Explainable Artificial Intelligence framework for root-cause analysis by integrating welding process parameters with microstructural failure modes. An experimental-computational design was applied to 120 AH36 marine-steel specimens produced by Gas Metal Arc Welding under controlled variations in current, voltage, travel speed, wire-feed rate, shielding-gas flow, and heat input. Weld quality was evaluated using visual inspection, ultrasonic testing, metallography, scanning electron microscopy, and Vickers microhardness, while Random Forest, XGBoost, and neural-network models were compared and interpreted using SHAP. Parameter variation enabled AI explanations to be checked against metallurgical evidence and defect morphology. XGBoost achieved the best performance, with 90.0% accuracy, a macro F1-score of 0.886, and an AUROC of 0.956. Heat input, shielding-gas flow, travel speed, and HAZ microhardness emerged as the dominant explanatory variables. Porosity was mainly associated with inadequate shielding, incomplete fusion and penetration with insufficient effective heat input, and cracking with elevated HAZ hardness. The study concludes that physically validated XAI can extend defect classification toward transparent, mechanism-informed diagnosis and provide a stronger basis for corrective decision-making in safety-critical marine welding.]]>
