Fatigue life prediction of additively manufactured alloys using a transformer framework optimized by hybrid leader-based algorithm

Estimating fatigue life of additively manufactured (AM) metals is challenging due to complex defects and variable loads. To address high experimental costs and limited generalization of conventional models, a Transformer–Hybrid Leader-Based Optimization (HLBO) framework is proposed for cross-material fatigue-life prediction. Eight features are extracted from 1,007 curated samples of four common alloys in FatigueData-AM2022 (originally 15,146 records of 116 alloys): temperature, stress ratio, frequency, stress-concentration factor, yield/ultimate tensile strengths, elongation, and nominal stress. The Transformer’s self-attention captures feature dependencies, while HLBO optimizes hyperparameters (e.g., lr=0.0005, 1 encoder/2 decoder layers). On the test set, the model achieves R 2 =0.90, MAE=0.18, MSE=0.06, and a composite index of 3.23, outperforming CNN, MLP, and ResNet baselines. The results indicate that the framework enhances prediction accuracy and generalizes strongly across materials and loading conditions, providing an efficient tool for reliability-driven design of AM components.