Data-Driven Prediction of Probabilistic <i>S-N</i> Curves for Steels Based Oncomposition and Processing Parameters

Knowledge of the probabilistic S-N (P-S-N) curves of structural materials is critical for their engineering applications. In the present work, a data-driven model to predict the P-S-N curves of steels is proposed with composition and processing parameters as inputs. The model is configured with Extreme learning machine, an emergent variant of Artificial neural network, to capture the nonlinearities in fatigue life modeling. Moreover, the fatigue limit and ultimate strength, which are explicitly formulated via Symbolic regression, are integrated as two informative input features . The training of the proposed model is conducted on plenty of synthetic data produced by the probabilistic Stüssi model, instead of on sparse experimental data, so as to ensure the training effectiveness. Extensive rotating-bending fatigue data covering 207 steels are collected for model validations, and the results demonstrate that the proposed model is accurate and stable for P-S-N curve prediction of steels. On this basis, three new steels with strong fatigue resistance are successfully discovered, providing beneficial guidance for the rational formulation of concerned compositions and processing parameters.