基于深度学习的三种地震波阻抗反演方法比较

神经网络结构的差异性导致深度学习效果不同。为此，在对比全卷积神经网络(FCN)、卷积循环神经网络(CRNN)和时域卷积神经网络(TCN)的三种网络结构的基础上，通过正演模型测试，对比、分析基于上述三种深度学习的地震波阻抗反演方法的精度和计算效率；然后通过实际资料应用进一步对比三种方法的效果。模型测试结果表明，基于TCN的波阻抗反演的计算效率和反演精度相对较高，基于TCN、FCN和CRNN的波阻抗反演用时分别为82、68和264s，皮尔逊相关系数分别为99.15%、97.84%和98.14%。实际资料应用表明，基于TCN的波阻抗反演结果与测井资料更加匹配。该结论可为智能地震波阻抗反演方法的优选提供参考。

Comparative analysis of three seismic impedance inversion methods based on deep learning

The difference in neural network structure leads to different deep learning effects. Hence, upon the comparison of the fully convolutional neural network (FCN), convolutional recurrent neural network (CRNN), and time-domain convolutional neural network (TCN), this study uses the forward model tests to comparatively analyze the accuracy and computational efficiency of seismic impedance inversion methods based on the above three deep learning methods. Moreover, the three methods are applied to actual data for further comparison. The experimental results show that the computational efficiency and accuracy of TCN-based wave impedance inversion are relatively high. For wave impedance inversion based on TCN, FCN, and CRNN, the inversion time is 82 s, 68 s, and 264 s, respectively, and the inversion accuracy is 99.15%, 97.84%, and 98.14%, respectively. The actual data application reveals that the results of TCN-based wave impedance inversion match better with the logging data. This conclusion can provide a reference for the optimization and selection of intelligent wave impedance inversion methods.