Hybrid deep learning-quantum optimization framework for keyhole defect mitigation in friction stir spot welding of AA2014-T6 and AA7075-T6

Keyhole defects represent a critical challenge in friction stir spot welding (FSSW) of dissimilar aluminum alloys, significantly compromising joint integrity and mechanical performance in aerospace applications. A novel hybrid deep learning-quantum optimization (DL-QO) framework is proposed for real-time prediction and mitigation of keyhole through adaptive process control in AA2014-T6/AA7075-T6 joints in lap configuration. The framework integrates a multi-head attention-based convolutional neural network (MHA-CNN) with quantum approximate optimization algorithm (QAOA). Experimental validation across nine test conditions achieved a maximum tensile-shear load of 5421 N at 1200 rpm, 20 mm/min feed rate, and 8 s dwell time, corresponding to a shear strength of 428.67 MPa and a weld efficiency of 92.13%. The MHA-CNN achieved 92.3% classification accuracy with an F 1 -score of 0.94 for severe defect detection. SHAP analysis identified rotational speed (SHAP = 0.342) and dwell time (SHAP = 0.287) as the most critical parameters. Quantum optimization contributed a 31.2% performance improvement over deep learning alone, achieving 78.4% keyhole defect reduction with real-time inference latency of 8.5 ms.