Quantitative analysis of metals Rb, Sr, Y and Zr are here described using certified standards of several soils by applying Artificial Neural Networks in Energy Dispersive X-Ray Fluorescence spectra. In this work the neural networks were used to model X-ray spectra. The four elements mentioned above show intense line overlapping which prevents the simultaneous determinations of them. The neural network used was the backpropagation type and its configuration was optimized. The data input in neural network consists in admitting parts of the spectra with each elemental characteristic spectral line, and the network output are the concentrations of all four elements. Two other methods for comparation of the results were used, the Partial Least Squares (PLS) and the Principal Components Regression (PCR), both are used frequently in spectra modeling. These three methods allowed the simultaneous determinations of the metals. The Root Mean Square Error (RMS) for the output were evaluated and a value of 0.112 was found for the Neural Network and 0.127 for the PLS and PCR.