Robust Deep Learning Seismic Velocity Inversion Over Varying Acquisitions

Abstract

Retrieving subsurface velocity information from recorded seismograms generally involves solving an inverse problem using various optimization methods. Deep learning has recently become an emerging alternative technique to provide solutions for such velocity inversion tasks. However, due to the lack of labeled data in field applications, supervised deep learning methods proposed for velocity model estimations often rely on training with synthetic modeled data. Applying these deep learning models to field data can be severely limited by differences between the configurations used in the synthetic training data and those for field survey. In this paper, we investigate the impacts of different acquisition geometries on deep learning velocity inversion and propose an acquisition-robust deep learning framework to improve model’s applicability by mitigating the impacts of different acquisition geometries. We train the proposed deep learning model with synthetic training data transformed to an alternative domain, e.g., wavenumber-time domain in this study. We demonstrate that deep learning models trained in this way are more robust and allow generalization over different acquisition configurations.