A data-driven multiscale model SCA-DNN for 3D woven composites based on the damage evolution genome database

Abstract

The current data-driven multiscale models are limited by the challenge of Neural Networks (NNs) in mapping the high-dimensional microscopic physical fields, making it impossible for them to reveal the microscopic damage evolution behavior. In the work, a data-driven multiscale model SCA-DNN based on the material damage evolution genome database is proposed for the meso-micro damage analysis of 3D woven composites (3DWCs) under the small strain and quasi-static loadings. In the model, the mesoscale problem is solved using the Self-consistent Clustering Analysis (SCA) method, and the microscale problem is solved in an equation-free manner using Deep Neural Network (DNN) models based on the material damage evolution genome database. The SCA method is utilized for reduced-order computation of the homogenized stress and the microscopic dimensionally acceptable damage evolution data of the microscopic representative volume elements (RVEs). 200,000 sets of data are included in the damage evolution genome database. The benchmark tests of 3DWC under four loading conditions are conducted. The SCA-DNN method demonstrates three capabilities: (1) predicting the stress–strain curves and the damage modes in agreement with the experiments, (2) predicting the damage evolution consistent with the SCA² solutions, (3) achieving an efficiency improvement of several times compared to the SCA² solutions.