基于深度学习的地层沉积正演模拟代理模型构建与应用  

Construction and Application of a Proxy Model for Stratigraphic Forward Modeling Based on Deep Learning

作　　者：刘彦锋 段太忠[1] 龚伟 廉培庆 张文彪[1] 黄渊 LIU YanFeng;DUAN TaiZhong;GONG Wei;LIAN PeiQing;ZHANG WenBiao;HUANG Yuan(Petroleum Exploration and Development Research Institute of SINOPEC,Beijing 102206,China;Beijing Normal University,Beijing 100091,China)

机构地区：[1]中国石化石油勘探开发研究院,北京102206 [2]北京师范大学,北京100091

出　　处：《沉积学报》2023年第3期791-803,共13页Acta Sedimentologica Sinica

基　　金：中国石化科技部项目(P21038-3,P20077kxjgz);中科院先导A类项目(XDA14010204)。

摘　　要：地层沉积正演模拟方法能更真实地刻画地下地质体的分布规律,比传统的基于地质统计学的建模更有优势,但是条件化难度大,使其应用在实际油藏地质建模时面临较大挑战。地层沉积反演模拟提高了该方法的实用性,基于模拟结果与观测数据定量比较,地层沉积反演模拟在全局优化算法的驱动下不断修正地层沉积正演模拟输入参数,使模拟结果与观测数据吻合度达到最佳。由于反演系统优化参数多,非线性极强,收敛难度大,需要的迭代次数多,单次正演模拟耗时长,导致该方法效率较低。地层沉积反演模拟和深度学习算法中的生成对抗网络相结合,提出了构建地层沉积正演模拟代理模型的方法。以大量的碳酸盐岩地层沉积模拟的合成模型作为样本,通过神经网络训练,形成地层正演模拟器的代理模型,然后将其代入地层沉积反演模拟系统,避免了直接地层正演模拟的长耗时,加快了反演模拟的收敛速度。通过巴哈马西缘碳酸盐岩地层沉积模拟验证了该方法的可行性,采用学习后的生成网络进行沉积反演模拟,反演效率得到了大幅提升。尽管本文展示的是二维实例,也有望扩展应用在三维模型上。Stratigraphic forward modeling(SFM) describes the subsurface geological bodies distribution more realistically than traditional geostatistical modeling,but it is difficult to condition and therefore challenging to apply to practical reservoir geological modeling.Inverse stratigraphic modeling(ISM) improves the practicability of the method.Based on a quantitative comparison between simulated results and observational data,ISM uses a global optimization algorithm to continuously modify the SFM input parameters to find a best fit between simulated and observed data.However,ISM tends to be inefficient and time-consuming,because it has many optimization parameters,strong nonlinearity and long time-consuming of single iteration,and requires a large number of iterations.In this study,ISM was combined with a deep learning algorithm called generative adversarial network(GAN),to construct a stratigraphic forward modeling proxy model.A proxy SFM model based on a large number of synthetic samples formed by neural network training is substituted into the ISM,avoiding the long time-consuming of single iteration,to accelerate the convergence speed of the inversion simulation.The feasibility of the method was verified by application to carbonate stratigraphy on the western margin of the Bahamas.Using the learnt network generated for sedimentation inversion simulation greatly accelerates convergence speed.Although this study focuses 2-D examples,it is expected that the method can be extended to 3-D models.

关 键 词：生成对抗神经网络 代理模型 深度学习 地层沉积过程正演模拟 

分 类 号：P628.3[天文地球—地质矿产勘探]