基于生成对抗网络（GANs）的条带状油气储层多条件约束建模方法研究

多点地质统计学方法在条带状油气储层三维地质建模中得到了很好的应用，但是基于训练图像的多点统计规律还是不能很满意地再现条带状储层的地质模式，或者只能同时利用一个条件数据（如地震）。拟将深度学习中的生成对抗网络方法与地质建模相结合，形成全新的三维地质建模思路，这不仅是人工智能与油气地质交叉学科研究的热点和前沿，同时也是成熟油田精细地质研究的重要技术手段之一。该申请拟以碎屑岩地区地表河道和碳酸盐岩地区地下暗河两类条带状储层为建模对象，研制可反映条带状储层地质模式的储层原型模型集和对应的条件数据集；将生成对抗网络方法与地质建模算法相结合，形成多条件约束储层模拟器的训练方法；最后，基于原型模型集和条件数据集，训练出储层模拟器，可直接将井点、地震等条件数据生成为吻合条带状储层地质模式及条件数据的储层三维地质模型。该课题首次提出多条件约束储层模拟器的概念及训练方法，将对未来地质建模方法产生深远的影响。

Multiple point statistic (MPS) methods have been well applied in the three-dimensional geological modeling of band-shape petroleum reservoirs, but the training image-based multi-point statistical information cannot reproduce geological patterns of the band-shape reservoirs very well, or cannot use multiple conditioning data such as the seismic data. In the proposed research, we combine Generative Adversarial Networks (GANs) method in Deep Learning with geological modeling algorithms to form a brand new modeling approach. This research is not only one of the hot topics and frontiers of the inter-discipline between Artificial Intelligence and petroleum geology, but also one of the important technical means for detailed geological research in mature oil fields. This proposal intends to take two types of band-shape reservoirs, namely surficial channel reservoirs in clastic area and underground channel reservoirs in carbonate area, as the modeling objects, and develops a set of prototype geological models reflecting geological patterns of the band-shape reservoirs, and corresponding conditioning data set. We combine GANs with geological molding algorithms to form a method for training a conditional reservoir simulator. Based on the prototype geological model set and the conditioning data set, a reservoir simulator is trained using the above method. This simulator can directly convert observed geological data (e.g, well facies and seismic data) into reservoir models that perfectly match the geological patterns of the band-shape reservoirs and the observed geological information. This research proposes the concept and training method of the conditional reservoir simulator for the first time, which will have very profound influence on the future development of geological modeling methods.