Abstract: Experiments in data analysis with a multilayer perceptron neural network (MPNN) in South America are shown. The approach used Regional Physical-space Statistical Analyis System (RPSAS) as reference for training the network. From 51 available observation stations, only data from 5 stations and at 850 mb were used to train the network. Only geopotential height at two pressure levels (850 mb and 500 mb) was analyzed. The set of examples was available for this experiment was small. Since the observation and analysis grids (RPSAS uses the grid of Eta atmospheric model) were not equal, the stations closest to the Eta model grid were used. The result was that analysis made by the trained network was satisfactorily close to the observed field at 850 mb.