Search for a command to run...
• A novel GRU-FCNN hybrid deep learning architecture is proposed. This dual-branch model integrates a gated recurrent unit (GRU) to capture sequential features from multiaxial load paths and a fully connected neural network (FCNN) to learn high-order static features. This synergistic approach enables a more comprehensive modeling of the fatigue damage process. • A prediction framework without reliance on material-specific or path-specific assumptions. By vectorizing multiaxial load paths, this framework enables fatigue life prediction without relying on predefined material constitutive relations or fatigue damage mechanisms. Consequently, the model delivers accurate predictions for diverse materials (e.g., titanium alloys, aluminum alloys, stainless steel) and complex loading types (e.g., proportional/non-proportional, bending/torsion). • Systematic hyperparameter optimization and cross-validation are performed. A grid search algorithm is employed to systematically optimize key model hyperparameters (e.g., learning rate, number of hidden layers, and epochs). The model's generalization capability and stability are rigorously assessed using 5-fold cross-validation, ensuring robust predictive performance. • Strong generalization and extrapolation capabilities are demonstrated. Across a dataset of seven materials and sixteen load paths, most predictions fall within a factor-of-two scatter band. Multi-material joint training further validates the model's ability to extrapolate to unseen materials and load paths, highlighting its significant potential for transfer learning. Multiaxial fatigue life prediction is critical for the safety assessment of engineering structures. Conventional models are often tailored to specific materials and loading paths, which limits their accuracy and broad applicability under complex multiaxial loading. To address this issue, we propose a hybrid deep learning model that combines a GRU network with a FCNN for fatigue life prediction. Multiaxial loading paths are encoded as time-series inputs, enabling the GRU to extract temporal features of the loading history, while the FCNN learns a nonlinear mapping from these features to fatigue life. Hyperparameters are tuned via grid search, and model performance is evaluated using five-fold cross-validation. The proposed approach is validated on seven representative materials, including TC4 titanium alloy, 7075-T651 aluminum alloy, and X5CrNi18-10 stainless steel, under a range of proportional and non-proportional loading paths. In most cases, predictions fall within a factor-of-two scatter band, demonstrating strong adaptability and generalization across materials and loading paths without material-specific assumptions. Moreover, training on a unified dataset covering all seven materials further demonstrates cross-material generalization, confirming the reliability of the GRU–FCNN model for multiaxial fatigue life prediction.