[论文解读] Data Conditioning for Subsurface Models with Single-Image Generative Adversarial Network (SinGAN)

Lei Liu, Eduardo Maldonado-Cruz+3 more|arXiv (Cornell University)|Apr 7, 2024

Seismic Imaging and Inversion Techniques被引用 5

一句话总结

本论文提出一个用于 SinGAN 基于地下模型数据条件化的最小接收标准工作流程，包括各种条件检查和地质概念条件评估。

ABSTRACT

The characterization of subsurface models relies on the accuracy of subsurface models which request integrating a large number of information across different sources through model conditioning, such as data conditioning and geological concepts conditioning. Conventional geostatistical models have a trade-off between honoring geological conditioning (i.e., qualitative geological concepts) and data conditioning (i.e., quantitative static data and dynamic data). To resolve this limit, generative AI methods, such as Generative adversarial network (GAN), have been widely applied for subsurface modeling due to their ability to reproduce complex geological patterns. However, the current practices of data conditioning in GANs conduct quality assessment through ocular inspection to check model plausibility or some preliminary quantitative analysis of the distribution of property of interests. We propose the generative AI realization minimum acceptance criteria for data conditioning, demonstrated with single image GAN. Our conditioning checks include static small-scale local and large-scale exhaustive data conditioning checks, local uncertainty, and spatial nonstationarity reproduction checks. We also check conditioning to geological concepts through multiscale spatial distribution, the number of connected geobodies, the spatial continuity check, and the model facies proportion reproduction check. Our proposed workflow provides guidance on the conditioning of deep learning methods for subsurface modeling and enhanced model conditioning checking essential for applying these models to support uncertainty characterization and decision making.

研究动机与目标

• 动机：在地下模型中平衡地质条件化与数据条件化的必要性。
• 为像 SinGAN 这样的生成模型的条件化开发一个最小接收标准框架。
• 提供一个工作流程，引导条件化检查以支持不确定性表征和决策制定。

提出的方法

• 引入覆盖静态小尺度数据、巨大规模的穷尽数据以及局部不确定性的条件化检查。
• 评估空间非平稳性再现和多尺度空间分布。
• 通过连通地体数量和空间连续性检查等指标评估地质概念的条件化。
• 定义模型相部比例再现的检查，以确保地质的合理性。

实验结果

研究问题

• RQ1 有效对 SinGAN 基于地下模型进行条件化所需的最小接收标准是什么？
• RQ2 条件化检查如何量化 SinGAN 生成实现中的局部和全局数据保真度？
• RQ3 地质概念如何在多尺度的 SinGAN 条件化中被整合和评估？
• RQ4 在使用生成式 AI 的地下建模中，哪种工作流最能支持不确定性表征和决策？

主要发现

• 提出一个条件化检查的工作流程，以引导地下 SinGAN 模型的数据条件化和质量评估。
• 检查覆盖静态本地数据、穷尽的大规模数据、局部不确定性以及空间非平稳性再现。
• 通过多尺度空间分布、连通地体、空间连续性和地层相比例再现来评估地质条件化。
• 该框架旨在为将深度学习模型应用于不确定性表征和决策提供基本的条件化检查。