A Fatigue Strength Prediction Method for Ferrous Alloys Based on the CNN-GRU Model With Bayesian Optimization

Abstract

Fatigue strength prediction is crucial for ensuring the structural integrity and durability of metallic components under cyclic loading. While traditional methods rely heavily on extensive experimental testing, recent advances in deep learning offer promising alternatives for ferrous alloys with unknown fatigue strength. This paper proposes a CNN-GRU model with Bayesian optimization (B-CNN-GRU)-based fatigue strength prediction method for ferrous alloys. The model integrates convolutional neural networks (CNN) for extracting local interaction patterns among neighboring descriptors and gated recurrent units (GRU) for capturing cross-feature sequential dependencies, enabling a comprehensive analysis of complex material descriptors. To enhance performance, Bayesian optimization is employed to automatically identify the optimal hyperparameters of the hybrid architecture. A dedicated dataset comprising 1753 samples with 20 features is constructed from a public materials database. Experimental results demonstrate that the proposed B-CNN-GRU model significantly outperforms other baseline models, including CNN, RNN, LSTM, GRU, and CNN-GRU, achieving superior performance with an R² of 0.9913 and substantially lower MAE, RMSE, and MAPE. Further application studies on 35 unseen ferrous alloys confirm the model's excellent generalization capability, stability, and high prediction accuracy, showcasing its strong potential for practical engineering applications.