Mixing the datasets from different air quality monitoring stations to obtain an only dataset that represents a specific air contaminant concentration times series has been considered a difficult task not suited for linear models. The air contaminants datasets mixing is a relevant topic because helps to the data analysts like epidemiologist to obtain a time series from a location (city, town, neighbor) and relate it to another event or phenomena (morbidity and mortality events). In this work, we proposed the basic Artificial Neural Network (ANN) architecture has a merging method and we use it to obtain the daily average concentration for Particulate Matters (PM10, PM2.5), and O3 air contaminants from ten data sets to Medellín (Colombia) between 2008 and 2016. The ANN is a connectionist system inspired by the biological neural network, its goal is to learn the nonlinearities from the data and it is tolerant to modeling datasets with noise, outliers, and missing values. However, like other traditional techniques, the ANN can suffer overfitting. To control it, we used the weight decay strategy as regularization method and backpropagation with cross-validation as the learning algorithm. The algorithms were implemented in R programming language. We compared the air contaminants ANN mixed with the traditional mixing methods (the mean and median of the data sets). The merging performance of the ANN was better than the traditional method. In conclusion, the ANN merging can capture the dynamics and variability of the air contaminants datasets better than the traditional method. In addition, ANN are a suitable method to mix data values from a different air quality monitoring stations.