The applicability of neural networks in modelling the growth of short fatigue cracks

This paper examines the use of a neural network to model the chaotic behaviour of the growth of short fatigue cracks which are characterized by a decreasing crack growth rate with increasing crack length. Fatigue crack growth is modelled in terms of the Hobson short fatigue crack growth law. The neural network is exclusively trained and tested on Hobson's experimental data of short fatigue cracks propagating in a 0.4% carbon steel. The empirical constants d, a and C of Hobson's growth law are determined from the neural network predictions and are found to be within the following approximate ranges 63 < d < 400 (µm), -0.27 < a < 0.08 and 1 × 10^-4 < C < 509 × 10^-4 with no proportional relationship observed between the constant C and the applied cyclic stress. It is shown that neural networks are a viable computational tool for modelling the chaotic behaviour of short fatigue crack growth.