Physics-informed machine learning model for mode I fatigue delamination growth in composite laminates under different load ratios

Abstract

Fatigue delamination growth (FDG) is the predominant damage mode in composite laminates, with the potential to compromise the integrity and reliability of composite structures. The prediction of delamination propagation during cyclic loadings is therefore of great importance in several industrial applications. The emerging machine learning (ML) provides a new research paradigm to characterize FDG behavior. Incorporating physical knowledge into ML promises reliable predictions with limited data volumes. A self-consistent physics-informed ML prediction framework, consisting of two connected physics-informed ML models, is proposed in the present study. The first ML model employs experimental data to predict the strain energy release rate (SERR) under different load ratios (R-ratios). The SERR predictions from the first ML model, as a function of the crack propagation length a-a₀, are utilized to train the second physics-informed ML model to estimate the fatigue crack growth rate da/dN under different R-ratios. The Bayesian optimization (BO) is adopted during the ML training to ensure that all hyperparameters of each ML model are self-optimizing, thus eliminating the need for manual tuning. After training, the model is able to predict FDG behavior under different R-ratios as a function of the SERR. The proposed physics-informed ML framework was found to be superior to non-physics-informed ML models, and exhibited reliable performance in terms of prediction accuracy, interpretability, generalization and extrapolation.