基于数据与模型联合驱动的波阻抗反演方法

井插值初始模型为基于模型的反演提供的低频信息往往不够准确，导致该模型驱动方法容易出现较大的波阻抗预测误差且建模效率较低.为缓解这些问题，本文利用数据驱动的深度学习反演更加擅长预测低频阻抗的优势，提出一种基于数据与模型联合驱动的波阻抗反演方法.该方法联合地震和测井等数据，先后开展数据驱动和模型驱动的波阻抗反演.首先，数据驱动部分使用井旁地震记录、测井导出的波阻抗曲线以及井插值低频阻抗曲线，搭建以双向门控递归单元为主要模块的波阻抗智能预测网络.其次，该网络预测的波阻抗的低频分量作为数据驱动初始模型，替代井插值初始模型而参与模型驱动部分.最后，模型驱动部分在地震数据匹配和数据驱动初始模型的共同约束下开展基于模型的反演，获得最终的波阻抗结果.合成数据和实际数据测试表明，本文方法相比于单一的数据驱动或模型驱动方法能获得更高分辨率和更高精度的波阻抗反演结果，从而为后续储层预测提供可靠的弹性参数分布.

Data-model jointly driven acoustic impedance inversion

The low-frequency components provided by the well-interpolated prior model for the model-based inversion is usually inaccurate, which usually results in large errors of predicted Acoustic Impedance(AI)and inferior modeling efficiency via the model-driven method. To alleviate these issues, this paper leverages the preponderance that the data-driven method represented by deep learning inversion can accurately estimate the low-frequency impedance, and investigates the data-model jointly driven AI inversion method. The proposed method combines seismic and well logging data to carry out data-driven and model-driven AI inversion successively. Firstly, the data-driven part utilizes several seismic records at the well locations, well-log derived AI curves, and well-interpolated low-frequency impedance curves to build an intelligent AI prediction network based on Bidirectional Gated Recursive Unit (Bi-GRU). Subsequently, the low-frequency components of estimated AI via the network are used as the data-driven prior model, which replaces the well-interpolated prior model and participates in the model-driven part. Finally, the model-driven part implements the model-based inversion under the joint constraints of seismic data matching and data-driven prior model to obtain the final AI results. Synthetic data and real data tests demonstrate that the proposed method can generate higher accuracy and higher resolution AI results compared with the data-driven or model-driven method. The precise AI results can provide reliable elastic parameter distribution for subsequent reservoir characterization.