Axial–Torsional Fatigue Life Prediction Using Fatigue Mixer: A Neural Network Integrating Loading History Sequences and Equivalent Amplitude Descriptors

Abstract

Fatigue‐life prediction under axial–torsional loading is challenging because loading paths vary widely and axial–shear interactions become highly non‐linear under non‐proportional loading. Fatigue Mixer is presented as a streamlined gated Mixer‐based framework tailored to axial–torsional fatigue‐life prediction from two‐channel axial–shear loading‐history sequences, with two equivalent‐amplitude descriptors incorporated through a late‐fusion head. The residual backbone alternates time‐ and channel‐mixing blocks, and multiplicative gating provides adaptive representation modulation; a weak monotonic regularization term provides an auxiliary consistency bias. The framework is evaluated under a unified preprocessing and data‐splitting recipe on a dataset spanning nine metallic materials and 20 axial–torsional loading paths, with detailed results reported on representative material subsets. Experiments include in‐domain testing, ablation, benchmarking against representative baselines, leave‐one‐path‐out unseen‐path assessment, and post hoc representation analysis. Predictions cluster around experimental values, with most cases within the 1.5 times error band and the remainder within the 2 times error band.