Domain knowledge-guided data-driven prestack seismic inversion using deep learning

The essential task of reservoir characterization is to predict elastic/petrophysical parameters or facies from observed seismic data and evaluate their uncertainty. Deep learning-based methods gain great popularity because of their powerful ability to obtain exact solutions for geophysical inverse problems. However, those deep learning methods that use seismic data as the only input lead to difficult training and unstable inversion results (i.e., transverse discontinuity or geologic unreliability). In such circumstances, it is beneficial if prior knowledge of the model domain can be incorporated into the deep learning framework. Therefore, we have developed prior-based loss functions in the learning step of the deep learning models to ensure that the predictions find low errors on the training sets and that their results are consistent with the known prior. In addition, the Monte Carlo dropout (MC-dropout) technique is introduced for the quantitative assessment of the uncertainty of the prediction results. We determine the effectiveness of our framework in the application of prestack seismic inversion, in which the initial model built from well-log interpolation is used to design the prior-based loss function. We first perform extensive experiments on the synthetic data and find that our method can yield more stable and reliable results compared with traditional methods. Combined with the transfer learning strategy, the application to real data further demonstrates that our deep learning framework obtains more reasonable inversion results with more horizontal continuity and greater geologic reliability than traditional approaches.