Seismic resolution enhancement using physics-assisted seismic deconvolution network and domain adaptation

Seismic deconvolution (SD) is an effective tool for enhancing vertical seismic resolution. Deep learning (DL) has recently gained significant popularity among geoscientists and is now widely applied in SD due to its exceptional computational efficiency and ability to model complex nonlinear relationships. In general, DL-based SD methods rely on high-quality synthetic and field data labels to achieve optimal performance. Unfortunately, acquiring sufficient field data labels remains a considerable challenge, often leading to unsatisfactory results. To address this limitation, we develop a physics-assisted workflow for SD based on DL and domain adaptation (DA), called PASDNet-DA. It comprises two main components: a physics-assisted forward module and a data-driven inverse module. The data-driven inverse module, generated by U-Net architecture, enhances computational efficiency, whereas the physics-assisted forward module, constructed based on traditional SD approaches, ensures data consistency and accurate results. Initially, we use the synthetic Marmousi data set to train our PASDNet. To adapt the model for field data, we use DA, a specific form of transfer learning, to create the PASDNet-DA. The DA process bridges the gap between synthetic and field data domains by calculating the Jeffrey divergence between synthetic and field data without requiring field labels. This approach effectively measures the similarity between these two data sets. Finally, we validate the effectiveness of our model using 3D field data. The results demonstrate that our PASDNet-DA outperforms three state-of-the-art methods in improving vertical seismic resolution, even without field labels. Furthermore, the 2D seismic profile results crossing three wells reveal that our model can better align with well logs than state-of-the-art methods, highlighting its practical effectiveness.