Finding new marketable mosquito repellents is a complex and time-consuming process that can be optimized via modelling. In this context, a SAR (Structure-Activity Relationship) model was designed from a set of 2171 molecules whose actual repellent activity against Aedes aegypti was available. Information-rich descriptors were used as input neurons of a three-layer perceptron (TLP) to compute the models. The most interesting classification model was a 20/6/2 TLP showing 94% and 89% accuracy on the training set and test set, respectively. A total of 57 other artificial neural network models based on the same architecture were also computed. This allowed us to consider all chemicals both as training and test set members in order to better interpret the results obtained with the selected model. Most of the wrong predictions were explainable. The 20/6/2 TLP model was then used for predicting the potential repellent activity of new molecules. Among them, two were successfully evaluated in vivo.