A Heterogeneous Multi-Modal Parallel Deep Learning Framework for Multiaxial Fatigue Life Prediction of Metals

Multiaxial fatigue damage remains a critical barrier to the reliable design of metallic engineering components, particularly under complex non-proportional loading paths. Conventional critical-plane and energy-based models rely heavily on empirical assumptions and struggle to capture intricate path-dependent damage evolution, leading to limited accuracy and generalizability. To overcome these challenges, we propose a novel data-driven framework that integrates heterogeneous information sources through a three-branch parallel neural architecture. Specifically, a deep neural network (DNN) extracts static material intrinsic properties; a bidirectional gated recurrent unit (BiGRU) enhanced with a multi-head self-attention mechanism captures long-term temporal dependencies and adaptively weights key stages in loading histories; and a one-dimensional convolutional neural network (1D-CNN) extracts local strain fluctuation patterns. The three streams are fused to produce a comprehensive damage representation for end-to-end life prediction. The output of the model is the base −10 logarithm of fatigue life, i.e., lg(Nf). Evaluated on a public multiaxial fatigue dataset comprising 40 metals and 1167 samples, the proposed model achieves a coefficient of determination (R2) of 0.9052, achieving promising performance on a fixed data split (R2 = 0.9052); ablation studies demonstrate the complementary nature of the three feature extraction branches, though 5-fold cross-validation suggests that the performance gain requires further validation with larger datasets. Systematic ablation studies confirm that the multi-source fusion strategy and the attention mechanism are pivotal to the performance gain. This work offers a robust and effective data-driven solution for high-accuracy fatigue life assessment, and points toward future integration of physical knowledge to further enhance extrapolation and transparency.