Seismic Velocity Model Building Based on the Multi-modal Learning

With the shift of hydrocarbon exploration targets toward deep and ultra-deep reservoirs, the integration of multi-source data, including well logs, vertical seismic profile (VSP) velocities, geological interpretations, and seismic data, has emerged as a pivotal approach for integrated velocity model building with seismic processing and interpretation. Nevertheless, persistent challenges still exist, i.e. insufficient reliability and automation in seismic interpretation, ill-posed inverse problem under diverse geological constraints, and the empirical selection of regularization parameters. To address these limitations, we propose a multi-modal learning-based velocity model building framework that integrates full-space data (e.g., initial velocity, migrated image, and relative geological time) with well-trajectory data (e.g., logging velocity). The former provides a low-resolution global background, while the latter delivers high-resolution local details. The proposed architecture comprises three core components. A generative adversarial network (GAN) processes spatial data to generate velocity models while ensuring consistency with geological interfaces. An auto-encoder extracts low-dimensional latent variables from logging velocity preserving velocity features. Adaptive instance normalization (AdaIN) modulates the GAN-derived feature maps using the latent variables from well logs, thereby aligning the output velocity distribution with the statistical characteristics of measured logging data. The proposed model-level multi-modal learning framework explicitly exploits interdependencies between spatial and borehole data, capitalizing on complementary information across modalities to achieve deep-learning-based velocity model building. Numerical experiments on the Marmousi II data demonstrate that the absence of any single modality degrades velocity accuracy. In field applications, compared to traditional interpretation-driven velocity model building, the proposed method generated a more reliable velocity model without user interaction and quality control.