This study established a machine learning framework to enhance the accuracy of very-high-cycle fatigue (VHCF) life prediction in selective laser melted Inconel 718 alloy by systematically comparing the use of generative adversarial networks (GANs) and variational auto-encoders (VAEs) for data augmentation. We quantified the influence of critical defect parameters (dimensions and stress amplitudes) extracted from fracture analyses on fatigue life and compared the performance of GANs versus VAEs in generating synthetic training data for three regression models (ANN, Random Forest, and SVR). The experimental fatigue data were augmented using both generative models, followed by hyperparameter optimization and rigorous validation against independent test sets. The results demonstrated that the GAN-generated data significantly improved the prediction metrics, with GAN-enhanced models achieving superior R² scores (0.91–0.97 vs. 0.86 ± 0.87) and lower MAEs (1.13–1.62% vs. 2.00–2.64%) compared to the VAE-based approaches. This work not only establishes GANs as a breakthrough tool for AM fatigue prediction but also provides a transferable methodology for data-driven modeling of defect-dominated failure mechanisms in advanced materials.

E36 steel, widely used in shipbuilding and offshore structures, offers moderate strength and excellent low-temperature toughness. However, its welded joints are highly susceptible to fatigue failure. Cracks typically initiate at weld toes or within the heat-affected zone (HAZ), severely limiting the fatigue life of fabricated components. Traditional life prediction methods are complex, inefficient, and lack accuracy. This study proposes a machine learning (ML) framework for efficient fatigue life prediction of E36 welded joints. Welded specimens using SQJ501 filler wire on prepared E36 steel established a dataset from 23 original fatigue test data points. The dataset was expanded via Z-parameter model fitting, with data scarcity addressed using SMOTE. Pearson correlation analysis validated data relationships. After grid-optimized training on the augmented data, models were evaluated on the original dataset. Results demonstrate that the machine learning models significantly outperformed the Z-parameter formula (R² = 0.643, MAPE = 16.15%). The artificial neural network (R² = 0.972, MAPE = 4.45%) delivered the best overall performance, while the random forest model exhibited high consistency between validation (R² = 0.888, MAPE = 6.34%) and testing sets (R² = 0.897), with its error being significantly lower than that of support vector regression.