Pre-Trained GAN Conditioning to Acoustic Impedance in Facies Modelling

This study explores an alternative approach to seismic conditioning for generative adversarial networks (GAN) in facies modelling. Instead of adversarial learning, the proposed training workflow adopts transfer learning to train a conditional generator, starting from a pre-trained unconditional generator. This reuses pre-trained neural network models, reducing the device cost of training a conditional GAN from scratch. By fine-tuning the configurations of this training workflow, we achieve a conditional generator that can create high-fidelity facies models conditioned to acoustic impedance maps.