Background. Atrial fibrillation (AF) is characterized by complex and irregular propagation patterns. One of the clinical goals in AF patients is to restore sinus rhythm, usually by ablation. Main targets of ablation are AF onset locations and drivers responsible for AF perpetuation. Aim. In this work, we propose to localize AF drivers from body surface potentials (BPS) using Deep Learning. Methods. We used realistic computerized models of atria (N=2039 nodes) and torso (M=659 nodes). Atria and torso models were used to simulate a normal sinus rhythm and 14 different AF propagation patterns in both left atria (LA) and right atria (RA), with different complexity and driver positions. We addressed the localization problem as a supervised classification problem. We divided the atria into 7 regions, as documented in the literature (see Fig. 1-(b)). Target variable, AF driver localization, represents seven classes corresponding to each atria region. AF drivers, if exist, were manually classified into one region for each time-instant. Input data are 64 BSPs for each time-instant with an SNR = 20dB (see Fig. 1-(a)). We trained a Multi Layer Perceptron with 3 hidden layers (100,50,50) units each. Output layer with 8 units, one per region, and one for no AF driver. We used ReLu activation function in hidden layers and Softmax in the output. Results We were able to correctly locate the 96% and 98% of drivers in the test and training sets, respectively (accuracy of 0.96 and 0.98), while the Cohen’s Kappa was 0.88 for both sets. Conclusions. The deep learning method we proposed can help to identify target regions for ablation using a non-invasive procedure as BSP mapping.