Unsupervised seismic acoustic impedance inversion based on generative diffusion model

Seismic acoustic impedance, defined as the product of rock density and seismic velocity, is essential for identifying different rock layers and their contents. Seismic acoustic impedance inversion (SAII) is a crucial technique for deriving high-resolution impedance profiles, aiding in the interpretation of subsurface geologic structures, reservoir identification, and lithologic characterization. Although deep-learning-based methods have become a new inversion paradigm, they often require high-quality labeled data and accurate low-frequency impedance models to produce high-resolution impedance profiles. We develop an unsupervised SAII method based on a generative diffusion model to address these limitations. Our approach begins by training the generative diffusion model using low-frequency impedance models as the conditional input, allowing it to capture the complex prior distribution from the training data. We then incorporate an explicit physical measurement model into the diffusion model sampling process to approximate the posterior distribution. Our method mitigates the dependency on low-frequency impedance and enhances inversion performance. Notably, our method is an unsupervised inversion framework, effectively addressing the inverse problem without the constraints of measurements and labeled data requirements. Synthetic and field data experiments validate our method, demonstrating superior accuracy compared with supervised deep learning, unsupervised deep learning, and regularization methods.