基于物理驱动的超深随钻方位电磁波测井深度学习反演

Physics-driven deep learning inversion for Azimuthal LWD electromagnetic wave measurement

超深随钻方位电磁波测井反演是表征地层参数信息的重要技术。基于正则化(物理驱动)的电磁波测井反演方法广泛应用于现场解释,但迭代过程中需要多次调用正演,计算耗时长且不能获得实时反演结果,因此迫切需要一种高效的反演方法对随钻电磁波测井资料进行实时反演。近些年,基于深度学习(数据驱动)的电磁波测井反演算法在油气勘探领域受到了广泛关注,但该算法过度依赖于数据本身,训练过程未考虑麦克斯韦理论,所以在数据集不完备的情况下,深度学习反演效果不佳。文中针对二维各向异性地层,提出了一种耦合物理驱动和数据驱动的混合反演流程:基于超深随钻方位电磁波测井数据,随机生成无断层和断层模型数据集进行网络训练;基于训练好的网络,实现模型预测。与传统深度学习方法相比,文中方法预测精度显著提高。对含有不同噪声水平的数据进行测试,结果表明:基于物理驱动的深度学习反演方法的电阻率模型反演效果良好,具有较强的鲁棒性,泛化能力更强。

Inversion for logging-while-drilling extra-deep azimuth electromagnetic measurement is an important technique to characterize formation parameter information.The inversion method for logging-while-drilling electromagnetic wave measurement based on regularization(physics-driven)is widely used in field interpretation,but it needs to utilize forward modeling for many times in the iterative process,which takes a long time to calculate and fails to obtain real-time inversion results.Therefore,an efficient inversion method is urgently needed for real-time inversion of logging-while-drilling electromagnetic data.In recent years,a deep learning(data driven)inversion algorithm of azimuthal logging-while-drilling electromagnetic wave measurement has attracted widespread attention in the field of oil and gas exploration,but the algorithm relies too much on data,overlooking the Maxwell theory during the training process.Consequently,the effect of deep learning inversion is not good when the data set is not complete.In this paper,a hybrid inversion method coupling physics-driven and data-driven methods is proposed for two-dimensional anisotropic formations:a network is trained using randomly generated datasets comprising models with and without faults,based on the data from logging-while drilling extra-deep azimuthal electromagnetic wave measurement;and model predictions are then made using the trained network.Compared with traditional deep learning methods,the prediction accuracy of the proposed method is significantly improved.Test results also show that,under the influence of different test noises,the physics-driven deep learning inversion method achieves favorable outcomes for resistivity models,exhibiting strong robustness and enhanced generalization ability.