基于晶体塑性力学框架的材料本构行为智能预测研究  

ARTIFICIAL INTELLIGENT PREDICTION METHODS OF MATERIAL CONSTITUTIVE BEHAVIOR BASED ON CRYSTAL PLASTICITY FRAMEWORK

作　　者：翁焕博 罗诚 袁荒 Weng Huanbo;Luo Cheng;Yuan Huang(School of Aerospace Engineering,Tsinghua University,Beijing 100084,China;Institute of Aero Engines,Tsinghua University,Beijing 100084,China)

机构地区：[1]清华大学航天航空学院,北京100084 [2]清华大学航空发动机研究院,北京100084

出　　处：《力学学报》2024年第12期3468-3483,共16页Chinese Journal of Theoretical and Applied Mechanics

基　　金：国家自然科学基金(92160204,12402071);国家重大研究计划(J2019-IV-0011-0079);中国博士后科学基金(2024M751635);国家资助博士后计划(GZB20240365);清华大学水木学者资助项目。

摘　　要：人工神经网络(ANNs)已逐渐成为非线性材料多尺度本构建模的重要工具.针对航空航天领域中广泛使用的镍基单晶合金开发了基于晶体塑性框架的材料本构行为智能预测方法.提出的新方法在数据驱动的基础上结合了晶体塑性本构模型,保留了晶体滑移系的求解框架,将激活滑移系上的状态变量作为网络的输入,建立了状态变量和滑移系剪切应变增量的物理联系,引入了物理信息损失函数,实现了应力的隐式求解,从而准确预测了单晶材料的单调、循环力学行为.进一步地,探究了不同损失函数对模型训练结果的影响,明确指出数据和物理约束共同作用下的模型性能显著提升.物理信息的融入在一定程度上提升了模型的外插预测精度,但在训练样本稀疏区域仍然无法做到精确预测.为了解决在训练样本稀疏区域难以精确预测的问题,在常规的离线学习策略上提出了在线学习策略,使得神经网络模型根据残差大小进行自学习,最终达到传统本构模型的预测精度.提出的基于神经网络的晶体塑性本构行为预测框架为材料本构关系研究领域提供了创新且有效的思路,有望进一步推动复杂材料的多尺度本构模型研究.Artificial neural networks(ANNs)have become a powerful and indispensable tool for multiscale constitutive modeling of nonlinear materials.This paper develops an intelligent prediction method for the constitutive behavior of nickel-based single-crystal alloys widely used in aerospace industries,based on the framework of crystal plasticity.The proposed approach integrates data-driven techniques with a conventional crystal plasticity constitutive model,resulting in a hybrid framework that enhances predictive capabilities while retaining the fundamental physical principles governing crystal slip systems.The solution framework preserves the conventional formulation for resolving slip systems,where state variables associated with these slip systems are used as inputs to the neural network.This enables the establishment of a physical relationship between the state variables and shear strain increments of the slip systems.A physics-informed loss function is employed to achieve the implicit stress integration,allowing the neural network to accurately predict the mechanical responses of single-crystal materials under both monotonic and cyclic loading conditions.The study also investigates the effects of different loss functions on the model training process,revealing that the combination of datadriven learning and physical constraints significantly improves the model’s performance.Integrating physical information within the loss function enhances the model's ability to extrapolate predictions beyond the range of training data,providing more accurate predictions for unseen scenarios.However,challenges remain in regions where training data is sparse,leading to less precise predictions.To address the limitations in sparse data regions,an innovative online training scheme is introduced on top of conventional offline learning strategies.This scheme enables the neural network to adaptively improve its performance by minimizing residual errors during predictions,essentially allowing the model to self-learn and refine its accura

关 键 词：循环晶体塑性 镍基单晶合金 物理信息神经网络(PINN) 取向敏感性 在线学习机制 

分 类 号：O346[理学—固体力学]