Deep Learning-Based Seismic Acquisition Pattern Design Guided by Velocity Model Imaging

The design of seismic acquisition patterns is a key challenge in geophysical exploration, as it depends on the strategic placement of sources and receivers to ensure sufficient energy interaction with target geological structures. State-of-the-art patterns, whether regular or irregular, are typically evaluated based on the quality of the resulting subsurface imaging. While regular patterns offer good coverage, they are often impractical due to environmental and economic constraints. Irregular patterns-such as random, jittered, or optimized designs-can suffer from noise sensitivity, aliasing, high computational costs, and limited adaptability to complex subsurface structures, as they usually rely on undersampled reconstruction frameworks to reduce the number of sources or receivers. This work proposes a deep learning approach to learn an optimized seismic pattern that reduces the number of receivers while preserving imaging quality. The method includes: (1) simulating shot gathers from a reference P-wave velocity model using the acoustic wave equation, (2) modeling the acquisition pattern as a trainable binary layer, and (3) guiding its training through a deep learning-based imaging operator. This ensures that reconstructed velocity models remain geologically and physically consistent with those from dense acquisitions. To enhance adaptability of the optimized pattern to variabilities in real-world geological settings, we propose using domain variability by perturbing the reference velocity model. Results show that the learned pattern outperforms uniform, random, and jittered designs in reconstructing P-wave velocity models under various undersampling rates.