Answers are grounded in this paper and cite their sources.
Demonstrates improved seismic inversion accuracy in small-sample conditions, suggesting effective adaptation strategies.
Seismic impedance inversion is crucial for oil and gas reservoir prediction. Conventional inversion methods usually suffer from severe non-uniqueness and low stability. Deep learning approaches exhibit remarkable capability to model complex nonlinear relationships and integrate multi-source data, thus possessing great application potential. However, field seismic data rarely contain sufficient well logs for model training, which easily leads to overfitting and poor generalisation of deep learning networks. To tackle this limitation, this study proposes a Meta-Learning Enhanced Generative Adversarial Network (ME-GAN) for small-sample seismic impedance inversion. By embedding a meta-learning mechanism into the generative adversarial network, the proposed framework enables rapid adaptation to new geological settings under limited sample conditions. Its core designs include three components: (i) a meta-learner module designed for fast knowledge learning from limited support sets to adapt to new geological settings; (ii) a meta-task generator that produces diverse learning tasks via spatial segmentation, data augmentation and frequency-domain transformation; (iii) a physical constraint module to guarantee the physical rationality of inversion results. The ME-GAN is implemented in three stages: pre-training, meta-learning-based fine-tuning and formal prediction. Comparative experiments on both synthetic and field data are conducted between the proposed ME-GAN and the conventional GAN. The results demonstrate that ME-GAN achieves substantially higher inversion accuracy than the conventional GAN under small-sample conditions, verifying that the embedded meta-learning mechanism effectively promotes rapid adaptation to new geological environments. The proposed method provides a feasible solution for seismic impedance inversion with limited well log samples and shows strong potential for practical applications.
No takes yet. Share an insight, caveat, or question.
Wang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: