A multivariate analysis using artificial neural networks for determining the relative contribution of the cationic composition of fluids (Na, K, Mg, Ca and Li) for the estimation of downhole temperatures of geothermal wells is here reported. Neural architectures were evaluated using different numerical techniques of training, activation function logistic and linear, several combinations of inputs, at most 20 neurons in the hidden layer and the measured temperatures as the targets. The obtained results in this paper shows that the relation log(Na/k) obtained the highest relative contribution {\color{black}(69% al 75%), whereas other variables such as, log(Mg/Na²) and log(Ca/Na²), showed a less contribution (3-13% and 12-22%, respectively). log(Na/Li), log(Li/√Mg) and Li obtained 3% variables had a relative contribution = 3%. The details of the methodology and the validation results are reported in this paper.