As a speedy analytical technique of chemical compositions, laserinduced breakdown spectroscopy (LIBS) is appealing in metallurgical industry for insitu, online or longrange applications. Combined with LIBS, neural networks are used to calibrate and quantify the concentration of Mn and Si of different kinds of steels. The performance of the neural networks with different inputs is studied. Compared with the common internal calibration methods, neural networks can utilize more information of spectra, and better correct the matrix effect and line interference. The inputs of the neural networks, however, need serious consideration, since they have a great effect on the measurement reproducibility and accuracy.