Modelling fatigue delamination growth in fibre-reinforced composites: Power-law equations or artificial neural networks?

Highlights: • Modelling fatigue-driven delamination requires cyclic and monotonic components of the square-rooted energy release rate. • Self-similarity principles inform the functional expression of the equations describing fatigue-driven delamination. • Single-hidden-layer neural networks adequately describe the effects of mode-mixity and stress-ratio on delamination growth. • Extreme machine learning allows representing physical constraints for modelling fatigue delamination growth. This paper discusses two alternative modelling approaches for describing fatigue delamination growth (FDG) in polymer-based fibre-reinforced composites, i.e. semi-empirical equations having a power-law form and artificial neural networks. Barenblatt's self-similarity principles are applied for identifying a suitable expression of the delamination driving force in terms of the square-rooted energy-release-rate range and the associated peak values. The general dependency of pre-factors and exponents of FDG power-laws on the stress-ratio and mode-mixity is discussed in detail. Single-hidden-layer neural networks (SHLNN) with the support of self-similarity principles are here proposed as an alternative to semi-empirical power laws for describing FDG in composites. A example application of SHLNN to mixed-mode and variable stress-ratio FDG is provided for the carbon/epoxy system T800H/#3631. The SHLNN predictions are compared to a semi-empirical fit based on a modified Hartman-Schijve power-law.