Application of neural networking for fatigue limit prediction of powder metallurgy steel parts

• An artificial neural network was utilized to predict endurance limits from carefully selected inputs.
• The worst-case correlation factor of 0.9 indicated that the neural network has been well trained.
• Comparison of predicted and experimental data confirmed the accuracy of the model.

A neural network was trained with existing fatigue strength data of unnotched PM steel samples fabricated under different experimental conditions. Samples had been tested with as-sintered or machined surfaces under three loading modes. The data were collected from published experimental investigations to predict the fatigue strength by an artificial neural network. Fabrication and testing parameters together with corresponding fatigue limit records were used as sets of data for network training. Network performance was established by its accurate predictions. Subsequently, a genetic algorithm was utilized to optimize experimental conditions, subject to practical limitations, to achieve desired fatigue strength values.