Abstract: The steel/CFRP interface bonding is determined by the melting width of CFRP in the direct steel/CFRP laser joining, while laser heat input is the major factor affecting the melting width. In this paper, the experiment of steel/CFRP direct laser joining was conducted and a BP neural network prediction model of the CFRP melting width was explored based on MLP and SVM, and the BP neural network was optimized by the AOA algorithm. Meanwhile, based on the RSM analysis, the weights of beam scanning rate, defocusing and laser power on the melting width were studied, and laser energy density distribution was calculated. The results indicate that the related coefficients R in the training, test and validation set of AOA-BP neural network are 0.96296, 0.97009, 0.98828, respectively, which means that the established AOA-BP prediction model exhibits well prediction accuracy and generalization capability. Furthermore, the results of RSM reveal that the descending order of the weights of laser power parameters influencing the melting width is as follows: beam scanning rate, defocusing, and laser power. The increase in defocusing serves to flatten the laser energy density distribution, expanding the coverage area of beam energy.