基于深度学习的二维斜率层析反演模型误差校正方法  

A deep learning-based method for error correction of 2D slope tomography-based inversion models

作　　者：葛大明[1] GE Da-Ming(Geophysical Research Institute,Shengli Oilfield Company,SINOPEC,Dongying 257022,China)

机构地区：[1]中国石化胜利油田分公司物探研究院,山东东营257022

出　　处：《物探与化探》2025年第2期385-393,共9页Geophysical and Geochemical Exploration

基　　金：中国石化重大科技项目(P22141)。

摘　　要：斜率层析成像是一种利用局部相干的地震反射波走时和斜率,反演地下介质宏观速度分布的方法,在地质构造复杂工区,斜率层析反演模型的误差较大。为此,本文提出一种基于深度学习的斜率层析反演模型误差校正方法。该方法以斜率层析反演模型作为神经网络输入,对应的理论模型作为标签,通过训练神经网络,建立从斜率层析反演模型到理论模型的非线性映射。为确保训练后的神经网络适用于实测地震资料,基于实测资料反演模型和偏移剖面生成训练样本。理论模型合成数据测试验证了所提方法的正确性和有效性。将该方法应用于滩浅海2D实测地震资料,获得了更高精度的速度模型和更高质量的深度偏移成像剖面。Slope tomography is a method to estimate subsurface velocity macromodels from the slopes and traveltimes of local coherent reflection events.In geologically complex areas,the macromodels obtained from slope tomography tend to yield larger errors.To address this issue,this study proposed a method for error correction of the models using deep learning.Specifically,with macromodels determined using slope tomography-based inversion serving as input and corresponding theoretical models as labels,a neural network was trained,yielding a nonlinear mapping from the slope tomography-derived macromodel to the corresponding theoretical model.To ensure that the trained neural network was applicable to measured seismic data,the training samples were generated from the inversion model and migration profiles of measured seismic data.Tests based on the data synthesized using the theoratical model validated the accuracy and effectiveness of the proposed method.The proposed method was then applied to the 2D measured seismic data from beaches and shallow seas,yielding velocity models with elevated precision and depth migration imaging profiles with high quality.

关 键 词：斜率层析成像 深度学习 误差校正 速度模型 神经网络 

分 类 号：P631.4[天文地球—地质矿产勘探]