Abstract:

Deep learning impedance inversion usually requires a large amount of labeled data for network training. However, in practical applications, labeled data (well logging data) is often limited. To address this issue, this paper proposes a new semi-supervised seismic inversion method based on Generative Adversarial Networks (GANs). The method improves the traditional GAN network structure by using a conditional GAN (cGAN), redesigning the generator with a Unet structure and the discriminator with a Resnet structure, and employing Wasserstein GAN (WGAN) to construct a new objective function. The network training is divided into two stages: first, training the discriminator with a small amount of labeled data, then, training the generator with a small amount of labeled data and a large amount of unlabeled data, where the generator is constrained by the forward convolutional model. Experimental results on synthetic data demonstrate that the proposed method is suitable for impedance inversion with limited labeled data, accurately reconstructs impedance models, and exhibits good noise resistance. Inversion results on real field data also indicate the practicality of this method. This method offers a novel approach for addressing the challenge of limited labeled data in seismic wave impedance inversion, holding promising prospects for practical applications.

Key words: deep learning, seismic wave impedance inversion, generative adversarial network, semi-supervised learning