![](https://congressarchive.cimne.com/composites_2023/assets/compobanner.jpg)
Prediction of Non-linear Response of Multiphase Composite Microstructures through Deep Learning of Reduced Structure-Response Data
Please login to view abstract download link
An approach to predict the non-linear response of a three-phase composite microstructure using convolutional neural networks (CNN) will be presented in this work. Mechanical properties of two-phase composite microstructures were predicted by deep learning of arbitrary microstructure images and power-law-based elastoplastic material responses. Finite element-based simulations were used to generate the in-plane mechanical response of arbitrary microstructures constituting the structure-property dataset. CNN models were then trained to accurately predict the properties of FRP microstructures as shown in Figure 1(b), with prediction errors of around 10%. It is known that manufacturing-induced voids significantly influence the mechanical properties of composites, especially in nonlinear regimes, whereby strength is more sensitive to defects than elastic properties. In this context, the present research aims to develop an interpretable deep learning technique to predict the non-linear mechanical response of FRP composite microstructures including manufacturing-induced voids. The structure-property data of random microstructure with voids are created using homogenisation-based FE simulations in Abaqus/Explicit, followed by appropriate dimensional reduction before being learned by the CNN. The proposed strategy can be used to design multi-phase microstructures and predict their non-linear stress-strain responses with better accuracy and minimal computational effort.