A combined machine learning and numerical approach for evaluating the uncertainty of 3D angle-interlock woven composites

The characterization and evaluation of uncertainties caused by the automated manufacturing procedure are essential for the early structural design and application of 3D woven composites. To account for the inherent uncertainties, numerical simulations must be integrated with statistical uncertainty quantification and propagation methods. However, in some cases, it is very challenging and computationally time consuming. In this paper, a multiscale uncertainty quantification method combining finite element analysis, wide & deep neural network and sensitivity analysis is proposed to probabilistically evaluate the tensile response of 3D angle-interlock woven composites. The required dataset for training and validating the model is created by a two-step numerical simulation. With the present model, a framework of the importance of each uncertainty in determining the macroscopic properties' variance can be established with fewer computational resources. The results indicate that the effect of uncertainties on the material tensile response is significant and the sensitivity information can serve as a guide for reducing uncertainties at the macroscale.

Automated summary

In this paper, a multiscale uncertainty quantification method combining finite element analysis, wide & deep neural network, and sensitivity analysis is proposed to probabilistically evaluate the tensile response of 3D angle-interlock woven composites. The required dataset for training and validating the model is created by a two-step numerical simulation. The results indicate that uncertainties have a significant impact on the material's tensile response and sensitivity information can guide the reduction of uncertainties at the macroscale.