材料塑性对于复杂形状制造和结构安全性至关重要，其研究对于深入理解材料性能、优化结构设计和提升材料性能具有重大意义。随着计算机技术的进步，基于晶体塑性的多尺度模拟方法已成为材料科学、机械工程和力学等领域的重要工具。晶体塑性有限元方法，结合位错滑移和孪晶机制，以及材料的晶体结构和取向信息，有效揭示了金属多晶材料的塑性变形机制。然而，当前基于最小能量原理和最大耗散功原理判断滑移系激活状态的方法存在局限性。

近年来，机器学习技术的快速发展为材料塑性模拟提供了新的视角,其强大的非线性映射能力为快速判断滑移系和孪晶系的激活状态提供了可能性。本文采用机器学习算法，成功建立了多晶材料宏观力学行为与微观位错滑移机制之间的联系。我们开发了一种神经网络模型，准确有效地确定了晶体塑性本构模型中的激活的滑移系和孪晶系，并将其应用于三种典型的多晶金属：面心立方（FCC）结构的高导无氧铜、体心立方（BCC）结构的α铁和密排六方（HCP）结构的AZ31B镁合金。在本文中，我们通过简单的神经网络模型，能够在较小的网络参数下准确预测滑移系和孪晶系的激活状态，并计算出相应的应变增量。与传统的晶体塑性程序相比，基于神经网络的晶体塑性方法，在求解奇异线性方程组方面更具有优势，能产生准确而有效的结果的同时，提升了计算效率。

Material plasticity is crucial for complex shape manufacturing and structural safety, and its research is of great significance for a deeper understanding of material properties, optimization of structural design, and improvement of material performance. With the advancement of computer technology, multi-scale simulation methods based on crystal plasticity have become important tools in fields such as materials science, mechanical engineering, and mechanics. The crystal plasticity finite element method, combined with dislocation slip and twinning mechanisms, as well as the crystal structure and orientation information of the material, effectively reveals the plastic deformation mechanism of metallic polycrystalline materials. However, the current methods for determining the activation status of slip systems based on the principles of minimum energy and maximum dissipated work have limitations.

In recent years, the rapid development of machine learning technology has provided a new perspective for material plasticity simulation. Its powerful nonlinear mapping ability provides the possibility for quickly determining the activation status of slip and twinning. This thesis uses machine learning algorithms to successfully establish the relationship between the macroscopic mechanical behavior of polycrystalline materials and the microscopic dislocation slip mechanism. We have developed a neural network model to accurately and effectively determine the activated slip and twin systems in the crystal plasticity constitutive model, and applied it to three common polycrystalline metals: oxygen-free high conductivity copper with face centered cubic (FCC) structure, α-iron with body centered cubic (BCC) structure and AZ31B magnesium alloy with hexagonal close-packed (HCP) structure. In this thesis, we accurately predict the activation status of slip and twinning systems with small network parameters through a simple neural network model, and calculate the corresponding strain increment. Compared with traditional crystal plasticity programs, neural network-based crystal plasticity methods have more advantages in solving singular linear equations, producing accurate and effective results while improving computational efficiency.