Machine learning-guided study of residual stress, distortion, and peak temperature in stainless steel laser welding

Abstract

This paper presents a machine learning (ML) approach using Kernel Ridge Regression (KRR) to predict peak temperature, residual stress, and distortion in stainless steels during oscillating laser welding. The model was trained using reliable data from numerical simulations, which incorporated both welding parameters and material properties of stainless steels. The KRR model’s regression analysis demonstrated high accuracy with R² values of 0.968, 0.951, and 0.928, and RMSE values of 3.35%, 4.51%, and 5.78% for peak temperature, maximum residual stress, and distortion degree, respectively. However, slight prediction deviations were observed, particularly at higher distortion levels. The study also highlighted the critical role of input feature weight functions in optimizing predictions. Peak temperature was predominantly influenced by physical material properties, while residual stress and distortion were governed by both mechanical and physical factors. Moreover, at lower peak temperatures, predictions were more sensitive to laser oscillation frequency, amplitude, and welding speed, whereas higher temperatures were more affected by preheating and sample thickness. Additionally, increased residual stress and distortion levels were strongly linked to the weight functions of laser oscillation frequency and amplitude.