Multiscale modelling with neural network-based crystal plasticity model from meso- to macroscale

Multiscale simulation plays a pivotal role in macroscopic behavior analysis by incorporating micro-level physical deformation evolutions, which provides valuable insights for materials science and industrial manufacturing. However, its development has remained relatively retarded despite the long-established concept due to the challenge of balancing efficiency and accuracy. To address this issue, the current study introduces a recurrent neural network-based constitutive model as the mesoscale surrogate for crystal plasticity, termed Polycrystalline Linearized Minimal State Cells (PolyLMSC), which comprises two parallel LMSCs to predict both the mechanical response and texture evolution simultaneously. A new texture-mechanics linkage method is proposed based on the Fourier coefficients of generalized spherical harmonic (GSH) functions, where the linearity of Fourier space promotes the PolyLMSC model to extend to different textures. To enhance generalizability across arbitrary loading conditions, arbitrary strain paths with random and diverse variations in incremental size and loading direction are adopted for data generation. During the validation at the single material point, the PolyLMSC model shows good generalization performance across different textures under arbitrary loading. Furthermore, the PolyLMSC model is evaluated through various component-scale simulation cases, illustrating reasonable accuracy at meso- and macroscale predictions with 1∼2 orders of magnitude improvement in computational efficiency compared to conventional crystal plasticity models. The validation results demonstrate the proposed model as a promising candidate for efficient and accurate multiscale simulations, bridging the meso- to macroscales.