Seismic Tomography Algorithm Using Neural Networks

The estimation of subsurface seismic velocities is fundamental in seismic data processing, as it enables the accurate positioning of reflections and diffractions within the subsurface. However, traditional methods often face challenges in scenarios characterized by strong velocity contrasts, blind zones, and complex geological structures, which limit the reliability of the resulting models. This study presents a seismic tomography approach based on neural networks. A total of 100,000 synthetic one-dimensional (1D) lithological models were randomly generated with increasing velocities and sharp contrasts. These models were spatially discretized into blocks, each assigned a specific velocity and density. Acoustic impedance and reflectivity profiles were derived from each model, and the reflectivity profiles were labeled according to their corresponding velocity models. The dataset was divided into training, validation, and testing subsets to develop and evaluate the model. The trained neural network accurately predicted velocity profiles from reflectivity data, even in cases involving strong velocity contrasts. The model exhibited strong generalization performance on unseen data, validating its robustness. This approach provides a fast and accurate alternative for estimating seismic velocity profiles, significantly reducing manual intervention and offering a reliable solution for subsurface characterization in complex geological environments.