As one of machine learning techniques, deep learning has recently achieved the state-of-the-art performances in many areas, such as computer vision, natural language processing, to name a few. A generative model called Generative Adversarial Network (GAN) was invented in 2014. This deep network model is deemed as the most interesting idea in the last 10 years by the machine learning community and outperformed the traditional methods in many tasks like image synthesis and super-resolution. Laying the heart of the GAN is its ability to model any realistically sharp data distribution. Instead of providing a "blurry" sample, the high-resolution samples can be sampled from the GAN model, no matter it is a natural image or a well log. In this abstract, we propose a novel ultrahigh resolution seismic inversion method using GAN priors. The basic workflow is described below. Firstly, a simple GAN architecture was designed. Then, we train this GAN to model the well-log data distribution. Once the GAN is properly trained, it offers the high-resolution samples as priors to the inversion algorithm. To effectively use this prior information, we adopt the projected gradient descent algorithm to iteratively fit the seismic data and projects the â€œblurryâ€ sample to the high resolution set of prior samples defined by the GAN. We further use a thin-layer model to validate the feasibility and superiority of our method. Comparing with the traditional method, our result shows a higher precision and resolution.
Presentation Date: Monday, September 16, 2019
Session Start Time: 1:50 PM
Presentation Time: 3:05 PM
Location: Poster Station 2
Presentation Type: Poster