Lithofacies Modeling Utilizing Generative Adversarial Networks and Geostatistics
Abstract
Artificial intelligence (AI) approaches have been introduced to water resources and geosciences as early as 1980s. In the last decade, AI is becoming an important research area to understand complex, nonlinear hydrogeologic systems owning to the availability of big data in geology and hydrology. However, failure to incorporate physics and the fundamental theory of matter could lead the AI models to be less trustworthy. This challenge of AI can be addressed by offering the scheme of hybrid analysis and modeling, which allows on physical models providing fundamental theory, and the uncertainty of physical models are processed with both interpretable and black box data-driven models. In this study, a hybrid of machine learning approaches and geostatistical approaches are employed to perform lithofacies modeling to thoroughly present geological characteristics. Hundreds of lithofacies realizations are simulated based on conditional probabilities from indicator kriging. In this case, a generative adversarial network (GAN), a minimax game between a generator and a discriminator, is utilized into 2D/3D lithofacies modeling with geostatistical realizations as training images. This research aims to introduce GAN into lithofacies modeling, which carries more geological complexity and computational cost. A case study of GAN-based lithofacies modeling is performed with 440 geophysical logs up to depth of 3000 ft in the East Baton Rouge Parish, Louisiana. The results shows that the GAN-generated lithological models honor both geological characteristics and the geophysical log data.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.H35M1171S