As an important metal material, steel is widely used in manufacturing and construction due to its good plasticity, toughness, and low price. China’s annual steel production and export volume ranks among the top in the world, and the scrap produced in the steel production process is an important resource. Accurate classification of steel scrap is a key link in electric furnace steelmaking, and it is also of great significance to the sustainable development of an environmental and energy. In order to improve the efficiency of scrap steel recycling, a method for intelligently identifying scrap steel grades using laser-induced breakdown spectroscopy (LIBS) combined with XGBSFS is proposed. Firstly, the LIBS data in the range of 170~400 nm were collected by Lapa-80 solid-state pulsed laser for 3 types of 18 different scrap steel samples. The gross error in the spectral data is eliminated by k-value verification, and the remaining data after the elimination is averaged. 28 groups for each sample, a total of 504 groups of average spectral data were obtained. Then, the spectral data is subjected to baseline correction, normalization and other preprocessing to reduce the influence of matrix fluctuations. Finally, the processed spectral data is divided into the training set and test set, and 16 characteristic spectral lines of Si, Cu, C and other elements in the spectral data are extracted as classification features for the model’s input. After the variables are optimized by the XGBSFS feature selection algorithm based on XGBoost, kNN and SVM are used to establish an intelligent identification model of scrap steel. The accuracy rates of the XGBSFS-SVM and XGBSFS-kNN algorithm models established in this study are 100% and 98.8% respectively, on the test set, and the input dimensions are also reduced from 16 dimensions to 2 dimensions, and the modeling times of the two models are respectively From 3.1 to 2.79 s, 3.26 s to 1.64 s. Compared with the algorithm model using SVM and kNN alone, the optimized model proposed in this paper has higher prediction accuracy, modeling efficiency, and better generalization ability. After comparing the comprehensive effects of modeling time and accuracy, the XGBSFS-SVM model was selected for the intelligent and rapid identification of different scraps. The experimental results show that the LIBS and XGBSFS feature optimization methods proposed in this study can effectively optimize the modeling of feature variables and provide a new technology for the rapid and intelligent identification of scrap steel types in industrial production and the recovery of steel.