Experimental qualification and physics-informed predictive modelling of robotic and manual MAG welding of S355MC steel

Abstract Robotic and manual MAG welding of 10-mm S355MC steel were compared using EN ISO 15614-1 qualification tests, sample-based capability estimates and physics-informed surrogate modelling. No non-conforming outcome was observed among 36 destructively tested robotic specimens (rule-of-three 95% upper bound, 8.3%), whereas five of 18 manual specimens were non-conforming (27.8%). All specimens came from single qualification weldments, so the counts characterise this campaign at specimen level only. Dispersion behaved the same way. Charpy impact-energy standard deviations reached 4.6–5.1 J in the two robotic modes but 19.5 J in manual welding; exploratory one-sided C pL estimates were 2.57, 2.54 and 0.71. One representative stress–strain record per process mode was reconstructed with a general regression neural network, keeping every fifth measured point (681–711 patterns), and the within-record leave-one-point-out RMSEs at h = 0.005 were 0.37, 0.44 and 0.48 MPa; transfer to new weldments remains untested. Descriptively, the Arc-Entropy Stability Index followed the Charpy coefficient of variation (Pearson r = 0.985, n = 3), and optical metallography on analogous welds agreed with, though could not confirm, the inferred phase interpretation. Campaign-specific and correlational: that is the status of every contrast and modelling relationship here, pending independent weldment-level replication.