<p>Fatigue life prediction of Mg alloy welded joints (MAWJ) is fundamentally constrained by scarce and highly dispersed high-cycle fatigue data, which limits the effectiveness of data-driven approaches. To address this issue, this study proposes a physics-constrained data augmentation strategy based on the Basquin–Goodman relations. Controlled perturbations are applied to the equivalent stress, and the corresponding fatigue lives are recalculated in a physically consistent manner to systematically expand the fatigue dataset. The results demonstrate that the proposed augmentation strategy significantly increases data coverage and density, effectively suppresses overfitting, and improves the predictive performance of machine learning models for MAWJ fatigue life. Compared with models trained on the original dataset, the augmented models exhibit substantial improvements in <i>R</i>^<i>2</i>, <i>MAE</i>, and <i>MSE</i>, while the prediction accuracy within the twofold error band increases from 92 to 100%, indicating enhanced stability and generalization capability. These improvements are particularly evident under small-sample high-cycle fatigue conditions of Mg alloy welded joints, demonstrating that the proposed physics-constrained augmentation framework provides a reliable approach for fatigue life prediction when experimental fatigue data are limited.</p> Graphical abstract <p></p>