Cross-Attention Enhanced Conditional Diffusion Model for Seismic Facies Analysis

Complex subsurface reservoirs, such as those for oil, gas, and groundwater, are crucial in geophysical exploration. High-precision seismic facies analysis is vital for characterizing these reservoirs with high lateral variability and small vertical thickness, which is essential for oil and gas exploration. Traditional deep learning methods for seismic lithology prediction have improved accuracy over manual interpretation but rely on assumptions about data distribution. Since seismic data is complex and the distribution is unknown, this poses a challenge for reservoir prediction. To address these issues, we propose a conditional diffusion model with cross-attention to learn seismic facies distribution. The model adds Gaussian noise to segmentation labels in the forward process and uses seismic data as conditional information to restore original data in the reverse process. The cross-attention module enhances conditional control by integrating semantic segmentation embeddings into the diffusion model, improving reservoir prediction accuracy.