Beyond data scarcity: achieving cross-wire generalization in single-pass gas metal arc welding through inter-material feature transfer learning

A Random Forest model is developed to predict single-pass GMAW bead width and height from current, voltage, travel speed, derived heat input/polynomial terms, and filler-wire chemistry using a database compiled from published cross-sections. Trained on four steel wires, the model achieves R² = 0.97 for both targets. With only four ER70S-6 samples added, the transferred model reaches R² = 0.98 (width) and 0.96 (height) without degrading accuracy on the original wires; the maximum absolute error is 1.21 mm. Against a Box–Behnken response-surface model built from 17 experiments, the proposed approach delivers comparable or lower errors with far fewer new tests. Interpretability analyses (feature importance, decision paths, PCA/t-SNE, and cosine similarity) highlight heat input as the dominant driver and explain the cross-wire generalization.