The XVIth “Journées de Géostatistique”
Discover the research work of our Ph.D. student Charlie Garayt on simulating 2D geological models combining a deep generative adversarial network (GAN) and a Bayesian approach.
Simulation of Complexly Correlated Spatial Data with Deep Generative Adversarial Networks: Application to 2D Structural Geomodeling
September 7, 2023 – 3:00 pm – 3:20 pm
Authors: Charlie GARAYT, Pierre-Marie Gibert, and Jean LANGANAY (Geovariances); Samy BLUSSEAU (Center for Mathematical Morphology (CMM), Mines Paris, PSL University); Nicolas DESASSIS and Thomas ROMARY (Centre for Geosciences and Geoengineering, Mines Paris, PSL University).
Simulating structural geological models is a challenging problem due to the complex spatial data correlations. Implicit methods like the potential field method, based on geostatistics, struggle to characterize all the uncertainties and produce realistic geological models. Through deep learning, artificial intelligence (AI) offers new opportunities to solve this problem. In this work, we propose to use a deep generative adversarial network (GAN) to generate unconditional geological models trained on a synthetic training dataset. The GAN is then used in a Bayesian context to find geological models that honor conditioning data, such as drillholes. This approach allows simulating different geological models associated with different geological contexts. Finally, the proposed method offers better management of uncertainties than the current geomodeling method.