Discriminator-Augmented Denoising Diffusion Probabilistic Models for Seismic Data

Deep learning methods have gained significant attention in seismic denoising due to their superior ability to extract weak signals from prestack data, which are often smoothed out by traditional techniques. While conventional convolutional neural networks can be trained in an end-to-end manner, they often fail to capture the underlying data distribution. Generative models are capable of reconstructing more realistic seismic signals by learning the distribution. The primary generative models include Variational Auto-Encoders (VAEs), Generative Adversarial Networks (GANs), and the more recently proposed Denoising Diffusion Probabilistic Models (DDPMs). VAEs offer stable training but tend to yield results of limited quality. GANs can produce high-quality outputs via an adversarial discriminator but suffer from unstable training. DDPM could provide a favorable balance between output quality and training stability. However, the supervised training paradigm relies on high-quality labeled data, which is often scarce in geophysical applications. This limitation frequently leads to constrained generalization ability. Consequently, there is significant signal leakage for strong reflection events when applied to unseen data from different work areas.

To address this, we propose a novel discriminator-constrained diffusion model. Our key innovation is the integration of a discriminator into the DDPM framework. This adversarial component provides a powerful constraint during the training process. The hybrid training objective combines the standard diffusion loss with the adversarial loss, guiding the model to preserve critical reflections while removing noise.

We validate our method through comprehensive experiments. On synthetic data containing various noise intensities. Our method has an improvement of 0.7 dB for noisy data with 1% Gaussian noise compared to standard DDPM. More importantly, in cross-field tests, the proposed method has an improvement of 2 dB. Visualizations of denoised sections and difference profiles confirm that our approach better preserves reflections.

In conclusion, the incorporation of adversarial training into the diffusion process offers a robust solution to the generalization challenge in deep learning-based seismic denoising. Our work demonstrates a promising pathway for applying advanced generative models to practical geophysical data with limited labels.