Physics-assisted transfer learning metamodels to predict bead geometry and carbon emission in laser butt welding

Laser butt welding (LBW) with high quality is widely sought-after, but results in non-negligible carbon emission (CE). However, predicting bead geometry and CE of LBW is important and challenging, especially when facing new scenarios. In this study, we develop the LBW platform with a CE data acquisition system, and propose a physics-assisted transfer learning (PTL) methodology to predict bead geometry and CE in LBW by leveraging physical knowledge and data of different scenarios. Experiments are designed using the optimal Latin hypercube sampling, and conducted to acquire bead geometry of bare-plate welding and butt welding. The physics-assisted dimensionless analysis is introduced to evaluate and filter the experimental data of bead geometry. Kriging (KRG) metamodel is trained to derive the relation between processing parameters and bead geometry, and is mapped to apply the data trend of bare-plate welding to butt welding. Radial basis function (RBF) is then used as the residual compensation to construct KRG-RBF metamodel. A transfer learning metamodel is obtained by combining KRG and KRG-RBF metamodels using optimized weight coefficients. Similarly, a PTL metamodel is constructed via mapping operation and residual compensation on the analytical formula to predict welding CE for a different scenario. The carbon efficiency assessment adopted can contribute to LBW with low-carbon and high-quality. Finally, cross-validation and supplementary experiments are conducted to evaluate the prediction accuracy of constructed metamodels. The results show that the proposed PTL methodology can identify outliers caused by scenario anomalies and achieve superior prediction accuracy.