Particle Swarm Optimization-Based Neural Network for Predicting Flow Stress at Elevated Temperature of Differently Aged AA6110 Aluminum Alloy
摘要
The flow curves of differently aged AA6110 aluminum alloy at various temperatures up to 300°C with a constant strain rate of 10−3 s−1 were studied. A neural network (NN) was employed to predict the flow stresses, using three input features: true strain, temperature, and aging condition, and producing one output: true stresses. The hyperparameters of the neural network include the training algorithms, the number of hidden layers, and the number of neurons in each hidden layer. To enhance model performance and lower computational costs, hyperparameters, specifically, the number of hidden layers, neurons, and training algorithm, were optimized using grid search, Pareto front, and particle swarm optimization (PSO). The top-performing neural network structure, featuring two hidden layers, achieved an RMSE of 0.102 and a MAPE of 0.027%, requiring 134 seconds for computation. A simplified model with a single hidden layer and only 11 neurons attained acceptable accuracy (RMSE = 0.824, MAPE = 0.257%) while significantly reducing computation time to 3.76 seconds. Compared to traditional grid search, PSO reduced the hyperparameter tuning time by as much as 94%. This approach illustrates the practical utility of PSO-optimized neural networks for predicting flow stress in aged aluminum alloys, with potential for integration into simulation frameworks for advanced forming processes.