基于深度学习的塑性变形理论建模研究进展  

Research progress on theoretical modeling of plastic deformation based on deep learning

作　　者：王新云[1] 唐学峰 余国卿 龙锦川 温红宁 邓磊[1] 金俊松[1] 龚攀[1] 张茂[1] WANG Xin-yun;TANG Xue-feng;YU Guo-qing;LONG Jin-chuan;WEN Hong-ning;DENG Lei;JIN Jun-song;GONG Pan;ZHANG Mao(State Key Laboratory of Materials Processing and Die&Mould Technology,Huazhong University of Science and Technology,Wuhan 430074,China)

机构地区：[1]华中科技大学材料成形与模具技术全国重点实验室,湖北武汉430074

出　　处：《塑性工程学报》2024年第4期92-116,共25页Journal of Plasticity Engineering

基　　金：国家自然科学基金资助项目(52105337,52090043);国家重点研发计划(2022YFB3706903)。

摘　　要：传统塑性理论模型由于引入了适当的假设和简化,面对复杂变形机制和复杂变形条件时具有局限性。基于深度学习的理论建模作为数据驱动的建模新范式,具有精度高、通用性强的特点,是塑性理论建模的重要方向。介绍了深度学习的建模方法,如传统神经网络模型、物理启发式神经网络模型和神经算子网络,分析了各个模型的特点。总结了近年来基于深度学习的塑性理论建模方法在本构关系、损伤断裂模型、微观组织模型和多尺度模型等方面应用的研究进展,并分析了提高模型精度、泛化能力和可解释性的方法。最后提出了基于深度学习的塑性理论建模方法面临的挑战和未来发展方向。Conventional models of plasticity theory,by virtue of integrating specific assumptions and simplifications,has limitations in the face of complex deformation mechanisms and conditions.As a new data-driven modeling paradigm,theoretical modeling based on deep learning has the characteristics of high precision and robust universality,which is an important direction of theoretical plasticity modeling.The modeling approaches of deep learning,such as traditional neural network models,physics-informed neural network models and neural operator networks were introduced,and the characteristics of each model were analyzed.The research progress in recent years on the ap-plication of plastic theory modeling methods based on deep learning in constitutive relationship,damage fracture model,microstructure model and multi-scale model was summarized.The methods to improve accuracy,generalization ability and interpretability of models were analyzed.Finally,the challenges and future development directions in theoretical plasticity modeling methods based on deep learning were proposed.

关 键 词：深度学习 塑性理论建模 本构关系 微观组织演变模型 损伤断裂模型 多尺度模型 

分 类 号：TG316.8[金属学及工艺—金属压力加工]