A physics-informed machine learning model for predicting fatigue delamination of composite laminates

It is widely accepted that fatigue delamination growth is one of the primary causes of failure in composite structures over extended periods of operation. Over the past few decades, various models have been developed to predict fatigue delamination in composite laminates. These models need many costly and time-consuming fatigue tests to account for multiple factors (e.g. material type, stacking sequence, loading history). Machine learning (ML) algorithms are widely used to improve predictions in various fields, including fatigue life and fatigue crack propagation. However, using ML without physical constraints can lead to inaccurate fatigue predictions. Physics-informed neural networks (PINNs) incorporate physical principles to improve ML model interpretability and prediction accuracy. Nevertheless, the use of physics-informed models for fatigue crack growth is less developed compared to fatigue life evaluation. The objective of this research is to propose a physics-informed machine learning strategy for developing a reliable model that accurately represents fatigue delamination growth. The study involved the following steps: (1) conducting fatigue experiments under mode I delamination loading with varying fiber bridging; (2) analyzing the experiments using the Paris law and performing regression analysis on the curve-fitting parameters related to crack propagation; (3) proposing a semi-empirical relation for fiber-bridged fatigue delamination; (4) constructing a PINN-based machine learning algorithm with dynamic weight parameters and a dynamic learning rate. The results clearly demonstrate that this physics-informed ML model outperforms non-physics-informed ML models and theoretical models in prediction accuracy.