摘要
超材料是具有特殊机械性能的工程结构材料,可以通过设计单胞结构,定制超材料的力学性能。提出一种基于多保真度神经网络的超材料力学性能预测方法,采用拉丁超立方采样与物理试验、有限元分析等方法,构建包含两个保真度数据集的初始数据库。基于低保真度数据集,训练获得低保真度神经网络。冻结低保真度神经网络的通用特征层,对特定特征层基于高保真度数据集进行重训练,获得多保真度神经网络。将待预测结构件作为多保真度神经网络的输入,多保真度神经网络的输出即为预测得到的超材料力学性能。研究表明,所提出的方法预测精度与效率显著优于传统方法,为超材料的优化设计奠定了基础。
Metamaterial is engineering structural material with special mechanical property,the mechanical property of metamaterial can be customized by designing single cell structure.A prediction method for the mechanical property of metamaterial based on multi-fidelity neural network was proposed.An initial database containing two fidelity datasets was constructed by means of Latin hypercube sampling,physical experiment and finite element analysis.Based on the low fidelity dataset,a low fidelity neural network was obtained by training.Multi-fidelity neural network was obtained by freezing the general feature layer of low fidelity neural network and retraining the specific feature layer based on high fidelity dataset.The structure to be predicted is taken as the input of multi-fidelity neural network,and the output of multi-fidelity neural network is the predicted mechanical property of metamaterial.The research shows that the prediction accuracy and efficiency of the proposed method are significantly higher than the traditional method,which lays a foundation for the optimal design of metamaterial.
作者
邱荣英
李钼石
刘钊
QiuRongyin;Li Mushi;Liu Zhao
出处
《机械制造》
2024年第4期32-37,共6页
Machinery
关键词
超材料
神经网络
力学性能
预测
Metamaterial
Neural Network
Mechanical Property
Prediction
