带有数据驱动波形矫正的全波形反演方法研究

The high-resolution velocity model (velocity model with high spatial resolution) is crucial in the exploration of the lithologic reservoir. Full waveform inversion (FWI), which is based on the observed seismic data, enables us to potentially build the high-resolution velocity model for the lithologic reservoir. However, the forward model (wave equation) of the existing FWI, which is used to generate the calculated data, can not accurately characterize the characteristics of the seismic wave propagation. As a result, FWI is inevitable to introduce artifacts into velocity model during the data fitting of the calculated data and the observed seismic data. To solve the above mentioned problem, we propose to set up FWI under the framework of combined data-driven and model-driven instead of directly using the framework of the existing model-driven FWI. Specifically, We firstly use the deep neural networks algorithm to correct the waveform information of the observed seismic data under the data-driven framework. Therefore, the waveform information determined by velocity is only preserved. Secondly, a new FWI method with the data-driven waveform correction mechanism is proposed. Meanwhile, a high efficient solving algorithm is suggested to the new FWI method. Based on the above study, we not only propose an effective method for building high-resolution velocity model but also provide a core technology for the exploration of the lithologic reservoir. As a consequence, this project will has not only a significant meaning on theory but also broad application prospects.

高精度速度模型（具有高空间分辨率的速度模型）对岩性油气藏勘探具有重要意义，基于地震观测数据的全波形反演具有为岩性油气藏构建高精度速度模型的潜力。然而，现有全波形反演生成计算数据时所基于的正演模型（波动方程）无法准确描述地震波的传播规律，导致以计算数据与地震观测数据波形匹配为目标的全波形反演不可避免的在反演速度模型中引入错误结构。为解决上述问题，本项目拟突破现有模型驱动框架，在数据驱动与模型驱动相结合的框架下提出新的全波形反演方法。首先，在数据驱动框架下基于深度神经网络算法对地震观测数据进行波形矫正，仅保留数据中由速度参数决定的波形信息。在此基础上，提出带有数据驱动波形矫正的全波形反演方法，并建立高效求解算法。本项目将提出反演高精度速度模型的有效方法，可为岩性油气藏地震波探测提供核心技术，具有重要的理论意义及广阔的应用前景。

地震勘探是岩性油气藏勘探的主要方法，声波速度是油气检测的主要参数，全波形反演是构建高分辨率声波速度参数的重要途径。然而现有全波形反演方法所依赖的正演模型（声波方程）无法准确描述岩性油气藏地震波的传播规律，导致全波形反演面临正演模型不准确这一理论问题。本项目突破传统模型驱动框架，提出在模型数据双驱动框架下构建新的全波形反演方法用于解决传统方法的理论不足。首先，针对深度学习算法进行了研究，提出了基于全局优化的模型驱动深度学习，为后续研究奠定了算法基础；其次，在数据驱动框架下提出了基于深度学习的地震观测数据波形矫正算子，实现了对地震数据中无法被声波方程所描述的数据成分进行波形矫正，为全波形反演建立了数据基础；最后，提出了模型特征域非线性全波形反演方法，为求解强非线性反问题提供了有效方法。经过上述研究，本项目为岩性油气藏地震勘探建立了新的声波速度高分辨反演方法，基本达到了既定的研究目标。研究成果为岩性油气藏地震探测提供核心技术，具有重要的理论意义和广阔的应用前景。