The content of manganese and nickel in the steel refining process will affect the hardness and brittleness of the final product, but the added content needs to be strictly controlled. At the same time, the traditional steel composition detection equipment had a high cost, low efficiency and slow speed. Therefore, a high-precision, fast and real-time analysis method is needed. This article used genetic partial least squares (GA-PLS) combined with LIBS technology to quantitatively detect the two elements of Mn and Ni in the spectrum of steel samples and compared the results with the quantitative analysisof traditional PLS to verify the predictive performance of the GA-PLS model. This experiment used 12 steel samples purchased in the steel market, the spectral information of 9 samples was used as the calibration set training model, and the spectral information of 3 samples was used as the test set to verify the quantitative performance. GA-PLS continuously raised the threshold of the selected frequency of the variable, established the PLS model with the variables under different thresholds, and compared the threshold when the lowest RMSECV was selected (the optimal thresholds for the selected frequency of the spectral input variables of Mn and Ni were 8 and 7 respectively). The results of GA-PLS showed that the $R_{P}^{2}$ and RMSE_P of the GA-PLS manganese prediction results were 0.999 0 and 1.347 3, and the relative analysis error (RPD) was 2.5; the $R_{P}^{2}$ and RMSE_P of the nickel prediction results were 0.999 5 and 0.525 4, respectively, and the RPD was 8.6. The final predicted result was better than PLS. The results show that the GA-PLS algorithm has the potential for sustainable mining in metallurgical metal element analysis, and will also promote the deeper application of LIBS technology in the field of steel smelting.