Abstract: A machine learning model for fatigue life prediction of TC4 welded joints is constructed based on very high-cycle fatigue tests and data augmentation using different interpolation methods. Very high cycle fatigue tests were conducted on TC4 welded joints under a stress ratio (R') of −1, and the original dataset characteristics, including stress amplitude, defect location, feature size and fatigue life, were obtained from S-N curves, SEM and Image J. Two machine learning models, namely back propagation (BP) neural network and random forest (RF), were established using the dataset augmented by nearest neighbor interpolation (NNI) and Gaussian mixture model-based nearest neighbor interpolation (GMM-NNI). The effects of different date augmentation methods on the predictive performance of these models were investigated. The results indicate that the failure mode for TC4 welded joints under very high-cycle fatigue is primarily internal failure, which is closely related to the fabrication process and stress ratio. The dataset augmented by GMM-NNI effectively improves the prediction performance of the machine learning models. The BP neural model, evaluated using specific metrics, demonstrates higher accuracy. When combined with data augmentation methods, the BP neural network model trained on the GMM-NNI dataset exhibits better generalization ability and prediction accuracy for the fatigue life of TC4 titanium alloy welded joints than the RF model.