基于机器学习嵌接有限元的蠕变时效成形全过程形性演变预测

Prediction of shape and property evolution in whole process of creep aging forming based on machine learning embedded into finite element

针对蠕变时效成形中存在的蠕变变形和时效强化动态交互耦合作用导致成形精度难以准确预测和控制问题,提出了一种基于机器学习的方法来预测蠕变时效过程中的形性演变。利用单向拉伸蠕变时效实验数据训练神经网络(NN)模型,用以描述蠕变时效本构关系。对比统一本构模型、反向传播NN(BPNN)模型、粒子群优化BPNN(PSO-BPNN)模型、遗传算法优化BPNN(GA-BPNN)模型对形性演变的预测效果,发现GA-BPNN和PSO-BPNN模型分别对蠕变应变和屈服强度具有较高的拟合精度。通过子程序将NN模型与有限元程序嵌接,实现了蠕变时效成形全过程的模拟,预测了铝合金板材蠕变变形和屈服强度的演变。针对回弹,相较于统一本构模型26.5%的误差,GA-BPNN模型的预测精度有较大提高,误差仅为5.1%。证明了采用机器学习的方法探寻蠕变时效本构关系并通过BPNN模型嵌接有限元模拟实现形性演变精确预测具有可行性。

Aiming at the problem of the difficulty for prediction and control of forming accurate due to the dynamic interaction between creep deformation and aging strengthening in creep aging forming,a method based on machine learning was presented to predict the evolution of shape and property during creep aging process.The data of uniaxial tensile creep aging tests was used to train the neural network(NN)model to describe the creep aging constitutive relationship.By comparing the prediction results of unified constitutive model,back propagation NN(BPNN)model,particle swarm optimization BPNN(PSO-BPNN)model and genetic algorithm optimization BPNN(GA-BPNN)model,it is found that GA-BPNN and PSO-BPNN models have higher fitting accuracy for creep strain and yield strength,respectively.The whole process of creep aging forming was simulated by embedding NN model into finite element program and the evolution of creep deformation and yield strength of aluminum alloy sheet was predicted.For springback,compared with the error of 26.5%of the unified constitutive model,the prediction accuracy of GA-BPNN model is greatly improved,and the error is only 5.1%.The feasibility of exploring the creep aging constitutive relationship by the machine learning method and realizing the accurate prediction of shape and property evolution through finite element simulation with embedded BPNN model was proved.