Machine Learning-Aided Cohesive Zone Modeling of Fatigue Delamination

Abstract The objective of this paper is to develop a machine learning-aided cohesive zone model (CZM) for fatigue delamination in composite structures. The so-called string-based CZM can handle pure and mixed fatigue delamination. Its solid thermodynamic foundation enables it to handle spectrum loading sequences well. An implicit integration scheme for this CZM is developed for improved accuracy and to generate needed training data. A conditional recurrent neural network (RNN) model can solve mixed sequential and time-invariant data problems. The time-invariant data are first input into a feed-forward neural network to predict the initial state of an RNN model. The RNN model will take the state and sequential data to recurrently predict the time series outputs. The conditional RNN model is trained to take the place of computationally costly finite element analysis (FEA) and then used for interface parameter calibration. The Dakota toolkit (a general-purpose optimizer), along with the trained conditional RNN model, can parameterize, automate, and accelerate model calibration. The trial-and-error process is then accomplished with Dakota and parameterized and automated with Python scripts and accelerate global optimum search with surrogate models. The present CZM is validated by calibrating its associated interface parameters from a series of constant amplitude double cantilever beam (DCB) tests on unidirectional E-glass fiber/E722 composite beams. It may be modified to accommodate other types of fracture or interfacial debonding.