摘要：

基于地震数据的岩性智能预测对于石油天然气的勘探开发具有重要意义。生成对抗网络(GANs)具有强大的非线性建模与模式迁移能力，已广泛应用于地质模型预测。然而，现有方法在处理复杂地质约束时，预测精度不足、结构连通性差、建模效率低，难以满足精细油藏描述需求。本文提出一种基于多尺度特征融合(MSFF)和并行卷积(PC)的生成对抗网络岩性智能预测方法。该方法通过设计多尺度特征融合架构，自适应提取地震特征，实现岩性的智能预测与不确定性量化，同时引入并行卷积模块以捕捉多尺度特征。结合多尺度特征融合与并行式分辨率生成策略，不断迭代提升智能预测可靠性。以南苏门答腊盆地LTAF组为例，从二维和三维数据测试分析验证方法的可靠性，并与多点统计方法(MPS)、支持向量机(SVM)和随机森林(RF)等传统预测方法进行对比分析。该方法能够快速实现对岩性的高分辨率预测，降低了预测结果的多解性，为下一步油气富集区的优选提供了指导依据。

关键词: 地质建模, 生成对抗网络, 多尺度特征融合, 并行卷积, 岩性预测

Abstract:

The intelligent prediction of lithology based on seismic data is of great significance for the exploration and development of oil and gas. Generative Adversarial Networks (GANs) have demonstrated strong nonlinear modeling capabilities and pattern transfer abilities in geological model prediction, but they still face some challenges in handling geological constraints and fine structures in terms of prediction accuracy, structural connectivity, and modeling efficiency. This paper proposes a lithology intelligent prediction method based on multi-scale feature fusion (MSFF) and parallel convolution (PC) of Generative Adversarial Networks. This method designs a multi-scale feature fusion architecture to adaptively extract seismic features, achieving intelligent prediction and uncertainty quantification of lithology, and introduces a parallel convolution module to capture multi-scale features. This study combines multi-scale feature fusion with a parallel resolution generation strategy to iteratively improve prediction reliability. Using the LTAF Formation of the South Sumatra Basin as an example, we verify the method’s robustness through 2D and 3D data tests and compare it with traditional methods such as Multiple-Point Statistics (MPS), Support Vector Machine (SVM), and Random Forest (RF). This method can rapidly achieve high-resolution prediction of lithology, reduces the multi-solution nature of the prediction results, and provides a guiding basis for the selection of potential oil and gas enrichment areas in the next step.

Key words: geological modeling, generative adversarial network, multi-scale feature fusion, parallel convolution, lithology prediction