Multi-fidelity data fusion for inelastic woven composites: Combining recurrent neural networks with transfer learning

Abstract

Surrogate deep learning models provide an efficient solution for reducing the computational demands of homogenizing complex meso-scale woven composites to study their elasto-plastic mechanical behaviors. This research introduces a comprehensive framework using transfer learning that combines data from a mean-field homogenization approach with high-fidelity full-field simulations. In a design space characterized by diverse loading conditions and micro-scale constitutive material properties, the goal is to address the challenges of generating sufficient datasets for training a data-hungry gated recurrent neural network (GRU). Multiple datasets of varying precision are generated and used, containing multi-axial stress–strain responses under two load types: random walking and proportional cyclic loading. Moreover, this study emphasizes the importance of temporal correlations in the dataset, which align with the physically path-dependent behavior of most non-linear materials, and demonstrates that temporal correlations are crucial for training time-series models. These correlations also provide the foundation for data augmentation using a linear interpolation technique within time-series stress analyses, enabling accurate predictions of homogenized meso-scale stresses based on strain trajectories and microstructural properties. Results demonstrate that integrating transfer learning with neural networks successfully incorporates a limited number of high-fidelity data with more accessible but low-fidelity data. With this framework, surrogate models for predicting the complex behavior of woven composites will be accurate and efficient, marking an important advancement in material modeling.