本团队博士生余国卿、唐学峰副教授等在Acta Materialia发表研究论文。

摘要：Data-driven surrogate model for crystal plasticity (CP) can significantly enhance computational efficiency and has the potential to advance the engineering applications of CP simulation. However, such surrogate model often lacks versatility and struggles to be applied effectively across different loading conditions and representative volume elements (RVEs), which limits its development and application. The purpose of this work is to develop a robust and scalable surrogate model for three-dimensional (3D) CP simulations under different loading modes and RVEs. To this end, a self-attention mechanism-based 3D convolutional neural network was developed to predict the full-field evolution of stress, strain, and crystallographic orientation. The predictive capability of the surrogate model and its scalability on RVE sizes, crystallographic textures and loading conditions were evaluated and discussed. The results indicated that the proposed surrogate model can accurately predict both the macroscopic and mesoscopic mechanical responses of RVEs with different grain and geometric sizes under various loading conditions. In terms of runtime, the proposed surrogate model accomplished the calculations in 2.81 seconds, which is approximately 2300 times faster than the 1.81 hours required by fast Fourier transform (FFT)-based CP simulation. The modeling framework proposed in this work significantly enhances the generalizability of the surrogate model for CP, providing a theoretical foundation for rapid cross-scale simulation under complex 3D deformation conditions and practical engineering applications.