Probability embedded failure prediction of unidirectional composites under biaxial loadings combining machine learning and micromechanical modelling

Lei Wan, Zahur Ullah*, Dongmin Yang*, Brian G. Falzon

*Corresponding author for this work

Research output: Contribution to journalArticlepeer-review

4 Citations (Scopus)
19 Downloads (Pure)

Abstract

This study presents a data-driven, probability embedded approach for the failure prediction of IM7/8552 unidirectional carbon fibre reinforced polymer (CFRP) composite materials under biaxial stress states based on micromechanical modelling and artificial neural networks (ANNs). High-fidelity 3D representative volume element (RVE) finite element models were used for the generation of failure points. Fibre failure and the friction between fibres and matrix after fibre/matrix debonding were taken into consideration and implemented as VUMAT subroutines, respectively. Uncertainty quantification was conducted based on a coupled experimental-numerical approach and failure probabilities were inserted into the failure points to generate the database for the training of ANNs. A total of 15 biaxial stress combinations were considered for the generation of datasets. Two strategies were considered for the construction of form-free failure criteria based on the ANNs for regression and classification problems. It is found that for the regression problems, an ANN model with 2 hidden layers and 64 neurons can achieve a mean square error (MSE) of 0.027% and a mean absolute error (MAE) of 0.78%. For the classification problems, an ANN model with 3 hidden layers and 32 neurons, presents an excellent performance in the prediction with a probability of 98.1%. A good agreement was observed between the failure strength of composites under transverse and in-plane shear predicted by these ANNs and failure envelopes theoretically predicted by Tsai–Wu and Hashin failure criteria.

Original languageEnglish
Article number116837
Number of pages17
JournalComposite Structures
Volume312
Early online date06 Mar 2023
DOIs
Publication statusPublished - 15 May 2023

Fingerprint

Dive into the research topics of 'Probability embedded failure prediction of unidirectional composites under biaxial loadings combining machine learning and micromechanical modelling'. Together they form a unique fingerprint.

Cite this