Deep Residual Conditional Generative Adversarial Network for Efficient Subsurface Modeling with Locally Varying Anisotropy

Transport in Porous Media · Springer Nature · 2026-11-16 · DOI: 10.1007/s11004-026-10354-1

Аннотация

Efficient modeling of subsurface properties with locally varying anisotropy is important for geological characterization, but conventional local-anisotropy simulation can be computationally demanding when large ensembles are required. A deep residual conditional generative adversarial network (cGAN) was developed to generate stochastic property fields from prescribed angle maps in a single forward pass. Training pairs combined prescribed angle maps for fold and fold-fault structures with standardized Gaussian fields generated by fast Fourier transform moving average (FFT-MA). Evaluation showed consistent orientation performance across independent cGAN training runs and when inference was extended beyond the training-window scale. Compared with direct sequential simulation with local anisotropy (DSS-LA) and FFT-MA, the cGAN reproduced the dominant orientation patterns, marginal distribution, and spatial continuity of the training targets, with slightly higher orientation errors and lower conditional dispersion. After training, the cGAN generated 1,000 realizations on a graphics processing unit (GPU) approximately 16 times faster than FFT-MA on the same GPU and 1,572 times faster than DSS-LA using 16 central processing unit (CPU) workers. Application to a post-stack seismic section further showed that orientations estimated from observed seismic data could condition the fixed cGAN to generate section-scale stochastic realizations of acoustic impedance without retraining. These results support the proposed cGAN as an efficient and scale-extensible complement to conventional simulation methods for two-dimensional subsurface modeling with locally varying anisotropy.

Читать у издателя →

Похожие статьи