Summary The manual detection of the first break (FB) in seismic data relies on visual recognition of amplitude and waveform variations by experts. However, for vast datasets generated during seismic acquisitions, this manual process becomes highly time-consuming and subject to individual interpretation. In this study, we propose an automated FB picking method based on neural networks for detecting the initial arrival of synthetic data resembling the characteristics of the Middle Magdalena Valley in Colombia, a basin historically associated with hydrocarbon exploration. Our approach involves supervised training of a neural network (NN) comprising a 1D convolutional layer and two dense layers. The NN categorizes reprocessed trace samples into two groups: pre-FB and post-FB. Subsequently, through post-processing, we identify the most likely sample corresponding to the FB. Our method successfully detects the first arrival in 77.36%, 91.19%, and 95.86% of cases, allowing for a margin of error of 10 samples when signal-to-noise ratios (SNRs) are 0 dB, 6 dB, and 20 dB, respectively. This demonstrates its effectiveness, particularly for noisy signal conditions.