From stress field to multiaxial fatigue life: A novel physics-guided neural network framework accounting for stress ratio, phase angle, and biaxiality ratio

• Physics-guided NN predicts multiaxial fatigue life from 3 principal-stress fields. • Unified model covers plain and notched specimens without scalar damage indices. • Principal-stress histories in the failure region are discretised in space and time. • Temporal Conv1D backbone learns phase- and path-dependent features from one cycle. Predicting the fatigue life under multiaxial loading is challenging because failure is governed by local stress fields and evolution of the stress state. The central idea of this study is to represent the geometric effect (including plain and notched configurations) by the spatial distributions of the three principal stresses in the potential region of failure, discretised over a set of nodes. Sampling these fields over one load cycle captures the loading history (stress ratio, biaxiality ratio, and phase angle) for constant-amplitude multiaxial loading, without introducing any additional ad hoc or phenomenological scalar damage parameter. To encode this loading history, a lightweight temporal Conv1D encoder is followed by a compact fully-connected regression head. The framework is validated on EN-GJS-600-3 using leave-one-case-out splits over 18 geometry–loading cases spanning multiple geometries, axial, torsional, and multiaxial loading, three stress ratios, two biaxiality ratios, and three phase angles. Across this broad set, the model achieves R 2 = 0.70, while adding the nominal metadata increases the prediction accuracy to R 2 = 0.75. Compared with classical critical-plane benchmarks (SWT and FS), the proposed framework achieves higher accuracy and remains applicable to both plain and notched specimens, whereas critical-plane criteria degrade strongly when plain data are included (due to differences in the dominant failure mechanisms). Sensitivity analyses confirmed robustness to reasonable choices of spatial and temporal discretisation. Gradient-based saliency illustrates mechanistic insights: spatial attribution concentrates at the notch tip, and temporal attribution is dominated by the first principal-stress, consistent with tensile crack-opening and with SWT, which outperforms the shear-driven FS criterion.