A variable-fidelity hybrid surrogate approach for quantifying uncertainties in the nonlinear response of braided composites
The ultimate strength prediction of textile composite materials requires high-fidelity FE modeling with information-passing multiscale schemes and damage initiation and propagation algorithms. The numerical demand of this procedure together with the complexity of the observed response surface, hampers the quantification of uncertainties contributing to the scatter of strength values. This study proposes a surrogate methodology able to efficiently emulate the nonlinear multiscale procedure, based on a combination of artificial neural networks and Kriging modeling under a variable-fidelity framework. A triaxially braided textile under longitudinal tension is used as a use-case and the methodology is employed to identify the most critical parameters in terms of variance via a global sensitivity analysis technique. Results show strong interaction effects between the uncertain parameters. The approach is non-intrusive and can be easily extended to other types of textiles and load cases.