Physics-driven cycle network for seismic impedance inversion using conditional generative adversarial networks

Abstract

Despite the extensive application of artificial neural networks in seismic inversion, their effectiveness is often hampered by the limited availability of labeled data. To address this challenge, we introduce a novel method for seismic impedance inversion. Our approach integrates a physics-driven cycle network with a Conditional Generative Adversarial Network (CGAN) and a convolutional model. Employing seismic data as input, the CGAN capitalizes on inherent information to minimize non-uniqueness during inversion. Furthermore, the convolutional model, acting as a physics-informed operator, reverts the derived impedance data back to seismic form, enabling simultaneous training of neural networks with labeled and unlabeled data, fulfilling the seismic-to-seismic cycle. The proposed method is demonstrated to be effective on tests using both theoretical models and field data.