Artificial neural network architecture optimization: comparing various training algorithms, transfer functions, and number of neurons using metal forming data

Cold rolling enhances metal properties through strain hardening; however, predicting the hardness of rolled samples remains challenging due to process complexity. Artificial neural networks (ANNs) can uncover hidden patterns among parameters to predict mechanical properties. However, determining an optimal ANN architecture is challenging. This study compares various ANN architectures for hardness prediction of 70-30 brass cold-rolled samples. A total of 2,028 architectures were designed and executed 202,800 times using a developed automated tool. Three ANN training algorithms, including Bayesian Regularization, Levenberg-Marquardt, and Quasi-Newton, combined with four transfer functions, LogSig, PureLin, TanSig, and ReLU, were evaluated on two-hidden-layer architectures with varying neuron counts. A dataset of 1,232 input-output pairs was used for training, validation, and testing. Results demonstrated that Levenberg-Marquardt delivered the best reliability, with high accuracy, fast convergence, and strong generalization to unseen data. In contrast, the Bayesian Regularization algorithm failed to generalize efficiently, making it unsuitable for this application.