Research on the prediction of mechanical properties of 316L stainless steel laser welded joints based on MLP

Abstract

To improve the efficiency and reliability of process design and parameter optimization in 316L stainless steel laser welding, this study develops a mechanical property prediction model based on a multi-layer perceptron (MLP) neural network. In traditional manufacturing practice, welding parameters are mainly determined through empirical experience and repeated trial-and-error experiments, which not only require substantial experimental workload but also result in long development cycles and high material consumption. This effect is more evident in thin plates and multi-thickness joints, where the interaction between process parameters and joint performance becomes more complicated. To address these challenges, key laser welding parameters, including laser power, welding speed, and plate thickness, are introduced as input variables to establish a data-driven model for predicting the ultimate tensile strength (UTS) of welded joints, providing a more efficient alternative to conventional experimental optimization. A dataset consisting of 400 welding samples with four different plate thickness configurations was obtained through controlled laser welding experiments to ensure data reliability. Before model training, Min–Max normalization and feature scaling were applied to reduce the influence of parameter magnitude differences and to improve numerical stability. In addition, a five-fold cross-validation strategy was employed to enhance the generalization capability of the model and to reduce overfitting. The results show that the proposed MLP model achieves an R² of 0.82 and a mean absolute error of approximately 54 MPa on the test set, indicating that the model can effectively capture the nonlinear relationships between welding parameters and joint mechanical properties. The prediction performance demonstrates the practical applicability of deep learning methods in laser welding process analysis and optimization.