Seismic inversion based on Markov chain Monte Carlo algorithms and attention mechanisms

Seismic impedance inversion using deep learning has been showing promising application prospects in reservoir characterization. However, the lack of labels of training datasets severely constrains the accuracy and stability of impedance estimation. In order to address these challenges, we developed a seismic inversion method based on Markov chain Monte Carlo (MCMC) algorithms and attention mechanisms in a supervised learning framework. Kernel density estimation (KDE) was employed to derive the underlying probability density function from well-log impedance data. A large number of impedance sequences with the target probability density distribution are generated using Metropolis-Hastings sampling of MCMC algorithms, and further used to augment training datasets containing synthetic seismic data and corresponding impedance labels. The architecture of the seismic impedance inversion model integrates deep convolutional neural networks (CNNs) with channel attention mechanisms, which enable the inversion model to selectively focus on critical features and improve inversion performance. We defined the physics-guided objective function consisting of seismic waveform loss and impedance loss terms. The application on field data demonstrates that the proposed method can enhance accuracy and resolution of seismic impedance inversion to characterize tight sandstone reservoirs within distributary channel systems.