Forward and Inverse Adversarial Model Applying to Geophysics

Abstract

The small sample problem limits the effectiveness of machine learning in geophysical logging, a crucial process in reservoir characterization. To address this, we propose a novel model combining generative adversarial networks with geophysical logging, called the Forward and Inverse Adversarial Model. Our contribution to artificial intelligence lies in the development of this model, which includes a forward component for generating synthetic logging data and an inverse component for predicting reservoir parameters. These components are trained using an adversarial process, enabling continuous improvement without relying on large labeled datasets. In the engineering application, our model enhances reservoir parameter prediction by generating high-quality logging data. Initially, the inverse model is pre-trained using measured data, where logging data serves as input and reservoir parameters as output. This model is then used to guide the forward model, which generates logging data based on virtual reservoir parameters. Both models are trained alternately, refining their performance until no further improvement is achieved. Experimental results on oilfield measured datasets validate the model's effectiveness in improving reservoir parameter predictions, particularly when the relationship between logging data and reservoir parameters is well-defined. The Forward and Inverse Adversarial Model demonstrates significant potential for improving geophysical logging data generation and subsequent reservoir analysis, advancing both artificial intelligent methodologies and practical engineering applications.

Keywords: Generative Artificial Intelligence； Well Logging； Forward Model；Inverse model；Machine Learning