Session: 01-06-01: Impact, Fatigue, Damage and Fracture of Composite Structures

Paper Number: 107351

107351 - Machine Learning-Aided Cohesive Zone Modelling of Fatigue Delamination 

In 2015, the Air Force Research Laboratory (AFRL) started a project to determine the technical feasibility of new damage tolerance design approaches for composite aircraft structures, and seven active research groups participated in this project. However, the average error for blind static strength predictions was 20%, while that for blind fatigue strength predictions was 40%. To realize the vision of composite service life prediction, we need to predict the remaining fatigue life of composite structures with fatigue test data. Cyclically loaded composite structures may fail by matrix fatigue and fatigue delamination. Existing models, however, cannot satisfactorily solve the complex problem. The objective of this paper is to develop a machine learning-aided cohesive zone model (CZM) for fatigue delamination in composite structures. Our previously developed CZM can handle pure and mixed fatigue delamination. Its solid thermodynamic foundation enables it to handle spectrum loading sequences well. We will first develop an implicit integration scheme for this CZM for improved accuracy and generate needed training data. Recurrent neural networks (RNNs) are a class of neural networks capable of recurrently solving time-dependent and sequential data problems. A conditional recurrent neural network (RNN) model reads the time series and time-invariant inputs and recurrently solves for the time series output. It can serve as a substitute for computationally costly finite element analysis (FEA) during model calibration. We will then train a conditional RNN model with all the inputs and outputs. Once trained, it can be used to calibrate the parameters of different interfaces. We will then calibrate the extended CZM as follows:
1.    Create an initial guess for the CZM.
2.    Perform structural analysis with the conditional RNN and the initial guess.
3.    Adjust the parameters until the predictions fit best to the test data (trial and error).
The Dakota toolkit (a general-purpose optimizer), along with a conditional RNN model, can parameterize, automate, and accelerate model calibration. We will therefore accomplish the trial-and-error process with Dakota. We will also parameterize and automate it with Python scripts and accelerate global optimum search with surrogate models. We will validate the machine learning-aided CZM as follows:
1.    Calibrate the interface parameters from a series of constant amplitude double cantilever beam (DCB) tests on unidirectional E-glass fiber/E722 composite beams.
2.    Compare the predictions by the conditional RNN model with those by FEA.
The resulting computational tool will achieve unprecedented predictive capabilities in fatigue delamination in composite structures. Such a tool will help reduce experiments and iterative adjustments, shorten the design and analysis period, and ultimately cut down the cost associated with developing and maintaining composite structures for the aerospace industry and other related industries.

Presenting Author: Liang Zhang AnalySwift LLC

Presenting Author Biography: Dr. Zhang is a Senior Research Scientist at AnalySwift, a university spinout company commercializing efficient and high-fidelity modeling tools for composites and other advanced materials and structures. He has been working in the field of composite structures for 10 years. His primary expertise is in nonlinear homogenization of composites, plasticity, damage, fracture, and fatigue. He has published 20+ journal articles and conference proceedings.

Authors:

Liang Zhang AnalySwift LLC
Xin Liu The University of Texas at Arlington
Su Tian AnalySwift LLC
Zhenyuan Gao Dassault Systèmes
Wenbin Yu Purdue University