Application of Kolmogorov–Arnold Networks for Predicting Flow Stress Curves of GH4698 Alloy in Hot Deformation

Kolmogorov–Arnold networks (KANs) have emerged as a significant advancement in neural network architectures, offering promising alternatives to traditional multi-layer perceptrons (MLPs). Unlike MLPs, which utilize fixed activation functions at their nodes, KANs incorporate learnable B-splines as activation functions along their connections, resulting in more adaptive network architectures. In this study, for the first time, we investigate the effectiveness of KANs in predicting flow stress curves of GH4698 alloy during hot deformation. To achieve this, hot compression tests were conducted on GH4698 alloy at various temperatures and strain rates, generating corresponding flow stress data. Based on the data, the traditional Arrhenius model, a MLP model, a KAN model with a similar architecture to the MLP model, and a simpler KAN model were applied for flow stress predictions. The prediction performance of four models was systematically compared. The results show that both the MLP and the KAN models exhibit higher accuracy than the Arrhenius model, demonstrating the advantages of machine learning methods in flow stress prediction. Notably, the KAN model outperforms the MLP model with a similar architecture in both prediction accuracy and convergence rate, and the simpler KAN model exhibits performance comparable to the MLP model. These findings indicate that KANs outperform MLPs in predicting flow stress curves of GH4698 alloy, highlighting their potential for enhancing predictive modeling.