基于深度学习的大地电磁二维反演研究  

Two-dimensional magnetotelluric inversion based on deep learning

作　　者：徐凯军[1] 卢炎 王大勇[2] 石双虎 XU Kaijun;LU Yan;WANG Dayong;SHI Shuanghu(School of Geosciences,China University of Petroleum(East China),Qingdao,Shandong 266580,China;Institute of Geophysical and Geochemical Exploration,Chinese Academy of Geological Sciences,Langfang,Hebei 065000,China;Acquisition Technology Institute,BGP Inc.,CNPC,Zhuozhou,Hebei 072751,China)

机构地区：[1]中国石油大学(华东)地球科学与技术学院,山东青岛266580 [2]中国地质科学院地球物理地球化学勘查研究所,河北廊坊065000 [3]东方地球物理公司采集技术中心,河北涿州072751

出　　处：《石油地球物理勘探》2024年第5期1174-1183,共10页Oil Geophysical Prospecting

基　　金：国家自然科学基金面上项目“基于结构和物性双重约束的井—地重力场和大地电磁场三维联合反演研究”(42274181)资助。

摘　　要：为了解决传统卷积神经网络反演由于层数过多而影响准确度的问题,开展了基于残差神经网络的大地电磁二维反演研究。通过大地电磁二维正演建立大量数据集,以TE和TM模式下的视电阻率和相位数据作为四通道网络输入,以对应的地电模型作为标签及输出进行有监督的学习,利用残差神经网络实现二维大地电磁反演。不同噪声水平的地电模型反演结果表明残差网络不仅可以很好地消除层数过多带来的准确度下降问题,还具有很强的抗噪作用。对冀中坳陷实测电磁资料进行反演,获得深部碳酸盐岩电阻率分布,据此分析了工区热储构造特征。理论模型和实测数据反演结果表明该方法具有良好的学习能力和抗噪性能,反演效果稳定可靠。This paper proposes two-dimensional magnetotelluric inversion based on residual neural networks to improve the accuracy of conventional inversion with convolutional neural networks,which are affected by excessive layers.A large number of data sets are established through two-dimensional forward modeling of magnetotelluric data.The apparent resistivity and phase data in TE and TM modes are used as input to a four-channel network,and the corresponding geoelectric model is used as a label and output for supervised learning.The twodimensional magnetotelluric inversion is achieved by utilizing residual neural networks.Based on the inversion results of geoelectric models with different noise levels,it is shown that residual networks can not only eliminate the problem of decreased accuracy caused by excessive layers but also exhibit strong noise resilience.Based on the inversion of the measured electromagnetic data in Jizhong depression,China,the resistivity distribution of deep carbonate rock is obtained,and the characteristics of the thermal storage structure in the working area are analyzed accordingly.Inversion results from theoretical models and measured data both show that the proposed method has excellent learning ability and noise resilience,and the inversion effect is stable and reliable.

关 键 词：电磁测深法 残差网络 深度学习反演 地热勘探 冀中坳陷 

分 类 号：P631[天文地球—地质矿产勘探]