摘要
广义有效介质极化理论(Generalized Effective Medium Theory for Induced Polarization,GEMTIP)提供了岩石物理参数与复电阻率(Complex Resistivity,CR)的频散关系,据此可基于观测到的激发极化(Induced Polarization,IP)数据反演岩矿石的激电参数。然而,传统的反演方法在非线性优化问题上存在局部最小值、计算量大和对初始模型依赖度高等问题,且含噪数据反演结果不稳定。此外,当前的激电参数反演研究主要集中在微观岩石孔隙表征和电化学机制领域,基于宏观地球物理观测数据直接进行反演和解释的相关研究不足。为此,提出了一种基于U-Net深度学习网络的方法,利用该方法可基于GEMTIP三维地电模型的地面IP差分数据直接提取激电参数。该方法将多个频率下的差分观测磁场振幅和相位作为网络输入,将异常区域的零频电阻率、体积分数、充电率、时间常数及弛豫常数作为输出标签。通过合成GEMTIP三维地电模型的可控源电磁样本数据训练深度神经网络,得到能够准确预测地下异常区域激电参数分布的网络模型。对包含GEMTIP激电参数的综合模型测试了该网络模型,并将测试结果与传统的正则化共轭梯度(Regularized Conjugate Gradients,RCG)反演结果进行比较,表明U-Net网络反演在耗时、求解精度和抗噪声能力方面均更具优势,能够从地面观测到的IP数据中直接、准确地预测GEMTIP激电参数。最后,利用深度学习方法对亚利桑那州南部North Sliver Bell地区的辉铜矿实际勘测数据进行训练,成功预测了该地区地下辉铜矿富集层分布,并与传统反演方法获得的地质解释成果进行对比,进一步证明了本文方法在实际应用中的可靠性和有效性。该方法可用于矿物组成和储层孔隙空间分布的预测,有望在宏观地球物理反演解释中得到广泛应用。
The generalized effective medium polarization(GEMTIP)theory provides the dispersion relation between rock physical parameters and complex resistivity(CR).According to this theory,the observed IP parameters can be used to invert the induced polarization parameter of rocks and ores.However,the conventional inversion method has some problems in nonlinear optimization,such as local minimum,a large amount of calculation,and high dependence on the initial model,and the inversion results of noisy data are unstable.In addition,the current research on IP parameter inversion mainly focuses on microscopic rock pore characterization and electrochemical mechanism,and the direct inversion and interpretation based on macroscopic geophysical observation data are insufficient.Therefore,a method based on a U-Net deep learning network is proposed.Based on this method,the IP parameters can be directly retrieved from the ground IP differential data of a GEMTIP three-dimensional geoelectric model.The method uses the amplitude and phase of the differential ob-servation magnetic field at multiple frequencies as network inputs,and the zero-frequency resistivity,volume fraction,charge ability,time constant,and relaxation constant of an abnormal region as output labels.The deep neural network is trained by the controlled source electromagnetic sample data of the synthesized GEMTIP three-dimensional geoelectric model,to obtain a network model that can accurately predict the distribution of IP parameters in underground abnormal areas.The network model is tested on a comprehensive model including GEMTIP IP parameters,and the results are compared with the traditional regularized conjugate gradients(RCG)inversion method.The test results show that the U-Net network inversion is superior to the traditional method in time,solution accuracy,and anti-noise ability,and can accurately predict the IP parameters of GEMTIP directly from the IP data observed on the ground.Finally,the deep learning method is used to train the actual exploration data of chalcocite in North Sliver Bell,southern Arizona,and the distribution of underground chalcocite enrichment layers in this area is successfully predicted.The results are compared with the geological interpretation data obtained through conventional inversion methods,which further proves the reliability and effectiveness of this method in practical application.The method can be used to predict mineral composition and spatial distribution of reservoir pores and is expected to be widely applied in macroscopic geophysical inversion interpretation.
作者
钟连诚
李伟勤
刘红岐
马琰祁
ZHONG Liancheng;LI Weiqin;LIU Hongqi;MA Yanqi(School of Electrical Information,Southwest Petroleum University,Chengdu,Sichuan 610599;School of Earth Science and Technology,Southwest Petroleum University,Chengdu,Sichuan 610599,China;State Key Laboratory of Oil and Gas Reservoir Geology and Development Engineering(Southwest Petroleum University),Chengdu,Sichuan 610599)
出处
《石油地球物理勘探》
EI
CSCD
北大核心
2024年第6期1433-1446,共14页
Oil Geophysical Prospecting
基金
国家自然科学基金课题“微纳孔密集岩石导电—介质机制与全谱检测理论研究”(41974117)资助。
关键词
广义有效介质模型
激发极化效应
激电参数反演
神经网络
岩性识别
generalized effective medium theory(GEMTIP)
induced polarization effect
induced polarization parameter inversion
neural network
lithological identificatio
