Implementation of denoising diffusion probability model for seismic interpretation

In this abstract, we show a novel machine learning-based diffusion model for seismic interpretation. In geophysics, reconstructing the subsurface structure from seismic data is an important inverse problem. Existing supervised machine learning (ML) solutions are to train a model to directly map measurements to seismic images, which are synthesized from images using a fixed velocity model. In this scenario, the generalization capability of models to the unknown measurement process could be hindered and out-of-distribution data could significantly reduce the inference accuracy from the pre-trained model. To address this issue, we implement the diffusion model, as a generative model, for the inverse interpretation problem and it provides a nature way to quantify uncertainty.