A comparison between transformers and recurrent neural networks for non linear elastoplastic behavior in 3D micro-mechanical modeling

During the last decades, Short Fiber Reinforced Composites (SFRCs) have gained attention in various industries for their favorable strength-to-density ratio, manufacturability, and costeffectiveness.They can be injection-molded into complex 3D shapes, giving rise to a wide variety of fiber orientation distributions, which necessitates the generation and homogenization of 3D Representative Volume Elements (RVEs), as illustrated in Figure 1.However, highfidelity micro-mechanical simulations of SFRCs remain computationally expensive -especially for non-linear, path-dependent elasto-plastic analyses, where each RVE homogenization involves a considerable numerical cost.Recurrent Neural Network (RNN) approaches have shown promise as surrogate models for history-dependent plasticity [1-2], yet challenges remain in accurately capturing long-term dependencies within a loading history.Recently, pretrained (PT) transformer models using self-attention mechanisms [2] have been proposed for surrogating micromechanical homogenization of 2D composites [4].By leveraging selfattention mechanisms, the transformer architecture can capture sequence-to-sequence dependencies and provides great accuracy for complex modeling cases.Yet, questions remain regarding potential specific advantages of PT transformers over RNNs for non-linear plasticity models, including handling data scarcity during training.