ConSeisDiff: A conditional diffusion approach to mitigate synthetic-real disparities in seismic fault detection

Abstract

Seismic fault detection is a critical task in geophysical exploration, demanding accuracy and efficiency in interpreting subsurface structures. While manual interpretation requires significant resources, deep learning models have become invaluable in automating the process of fault detection. However, the availability of labeled seismic fault data remains extremely limited, which pushes researchers to rely on synthetic images. In this work, we introduce ConSeisDiff (Conditional Seismic Diffusion) network, a novel conditional denoising diffusion model designed to generate synthetic seismic data for fault detection. Unlike traditional methods that rely on simplistic and conventional fault generation approaches, ConSeisDiff generates 2D seismic images by conditioning on fault layer attributes and facies maps. A key finding of our model is the introduction of a seismic conditional encoder, which incorporates seismic layer information extracted from the Canny edge algorithm, thereby enhancing the model’s ability to capture complex geological layers and structures. ConSeisDiff leverages a dual-path encoder–decoder architecture, combining residual blocks with transformer-based attention mechanisms to capture both local and global seismic features. We evaluate ConSeisDiff using several metrics, including MSE, PSNR, DSSIM, and FID, demonstrating that it outperforms state-of-the-art generative models in terms of fidelity and structural quality. Furthermore, we show that models trained on synthetic data generated by ConSeisDiff achieve performance comparable to those trained on real seismic data, effectively bridging the gap between synthetic and real-world applications.