Summary In seismology, energy is produced through the interaction between geological layers. This energy is released when some of the layers fail, producing a wavefield that travels to the surface and is recorded by seismological stations as traces. These traces are used to determine the magnitude and location of an earthquake, and to create a tomographic image (TI) of the Earth's interior. The TI allows for identification of the number of geological layers and their physical characteristics (e.g., background velocity) from the first break (FB) information. This work proposes a conditional Generative Adversarial Network (cGAN) to predict the FB in seismological data. The cGAN model is trained on the STanford EArthquake Dataset, STEAD ( Mousavi et al. 2019 ), which shares similarities with data from the Middle Magdalena Valley (MMV) in Colombia. The MMV is a region of geophysical interest due to the presence of oil and its proximity to the Bucaramanga Seismic Nest (BSN). After training and validation, the cGAN predicts FB on real earthquakes from the MMV. Overall, the proposed cGAN shows potential for improving seismic analysis in regions like the MMV, where traditional seismic methods may face challenges due to complex geological structures.