In the first place, we have developed a methodology for Automatic First Break Picking (FBP).

In geophysics, FBP is the process of identifying the first arrival signals within seismic traces. These signals are mainly associated with the energy of refracted waves at the base of the weathering layer or the direct wave traveling from the source to the receiver. The creation of a reliable and automated FBP algorithm is strongly needed, particularly when manual FBP struggles to efficiently handle large and low-quality datasets.

There are several well-known approaches to solve FBP. For example, Short-Term Average over Long-Term Average (STA/LTA) algorithm shows acceptable results, however, the STA/LTA and similar energy ratio-based methods are unstable for data with low signal-to-noise ratio (SNR) and require manual adjustment and conditioning, which is a time-consuming and resource-intensive process.

To address these limitations, we propose an advanced automatic FBP solution using deep neural network architecture. The model is trained on synthetic seismic data specifically designed to mimic target raw data. This step can be easily implemented with minimal parameterization and without understanding of the data specifics, which eliminates the need for manual efforts. First breaks can be obtained with high accuracy and efficiently by applying the trained model to entire seismic gathers.

Initial results are indicating significant potential for improving FBP accuracy and processing speed. The model demonstrates robustness to low SNR data, outliers handling, and adaptability to various first break curve behaviors. These advantages make our deep learning-driven FBP solution a valuable method for the seismic data processing cycle.