波动方程正演引导的深度学习地震波形反演  

Deep learning seismic waveform inversion based on the forward modeling guidance of wave equation

作　　者：段友祥[1] 崔乐乐 孙歧峰[1] 杜启振[2] DUAN Youxiang;CUI Lele;SUN Qifeng;DU Qizhen(College of Computer Science and Technology,China University of Petroleum(East China),Qingdao,Shandong266580,China;School of Earth Geosciences,China University of Petroleum(East China),Qingdao,Shandong266580,China)

机构地区：[1]中国石油大学(华东)计算机科学与技术学院,山东青岛266580 [2]中国石油大学(华东)地球科学与技术学院,山东青岛266580

出　　处：《石油地球物理勘探》2023年第3期485-494,共10页Oil Geophysical Prospecting

基　　金：国家自然科学基金项目“弹性全波形反演与逆时偏移成像”(41930429);中央高校基本科研业务费专项资金项目“面向深层油藏精细描述的地质知识库研究与开发”(20CX05017A);中石油重大科技项目“面向深层油藏精细描述的地质知识库研究与开发”(ZD2019-183-006)联合资助。

摘　　要：物理驱动的全波形反演方法计算成本高,数据驱动的深度学习反演方法对标记数据集的依赖性强。为了在有限的数据条件下获得更好的反演结果,结合数据驱动与物理驱动,提出了波动方程正演引导的深度学习地震波形反演方法。首先,利用地震数据应用神经网络重建速度模型,对网络预测的速度模型进行正演建模,通过最小化速度模型的误差及地震数据的误差训练网络;其次,使用有限差分法将二阶偏微分波动方程近似为可微算子,使正演过程能够传递梯度,并根据梯度方向动态调整地震数据损失的权重。实验结果表明,该方法能在一定程度上降低数据驱动方法对标记数据集的依赖性,可得到更准确的速度模型,且具有较强的鲁棒性。Physics-driven full-waveform inversion method has high computational overhead,while data-driven deep learning inversion method has a high dependence on the marker dataset.In order to obtain better inversion results with limited data,a deep learning seismic waveform inversion method based on the forward modeling guidance of the wave equation is proposed,which integrates data-driven and physics-driven methods.Firstly,a neural network is applied to reconstruct the velocity model based on seismic data,and the velocity model predicted by the network is modeled by forward simulation.Then the network is trained by minimizing the error of the velocity model and seismic data.Secondly,a finite difference method is adopted to approximate the second-order partial differential wave equation as a differentiable operator.The forward modeling process enables gradient transfer,and the weights of seismic data loss are dynamically adjusted according to the direction of the gradients.The experimental results show that the method can reduce the dependence of the data-driven method on the marker dataset to a certain extent,obtain more accurate velocity models,and have stronger robustness.

关 键 词：深度学习 地震反演 速度模型建立 正演建模 波动方程 

分 类 号：P631[天文地球—地质矿产勘探]