A supervised Machine Learning approach to classify traffic-derived PM sources based on their magnetic properties.

This study was intended to make a feasibility state about the discriminating power of magnetic parameters for a source-to-sink purpose by using a Machine Learning approach. The model appears to be very promising. Indeed, the results appeared very consistent giving the sampling locations of the predicted samples. This is fascinating that the discriminating power of source magnetic properties popped out from the model. The Machine Learning approach trained with the magnetic properties was able to differentiate origin signal of stations located near the one from the other, as we can see for the passive filters deployed in the street canyon. Nevertheless, the data base is finally too small. Since the samples are very heterogenous, there is a really need to integrate more and more new sample in order to increase the accuracy of the model. In another way, the model can also be improved, particularly by finding a way to include Na. For now, the model is not able to deal with a lack of data, this is unfortunate because for many reasons it is sometimes not possible to perform all the magnetic techniques on the same sample. In this case, these samples must be removed from the data set while more data we have to train the algorithm, better the model is. Another improvement that should be providing is the prediction of more than one category per sample to finally be able to quantify them. In that way, we will not predict the main source but we will estimate the contribution of all the defined sources involved in the measured magnetic signal. Finally, to really evaluate the efficiency of Machine Learning classification based on magnetic properties, it should be very interesting to compare the results with other source-to-sink modeling and isotopic chemistry analysis since both techniques are commonly used to discriminate the source signals