深度学习在地震速度建模中的应用研究综述  

A Review of the Application of Deep Learning in Seismic Velocity Modeling

作　　者：曹来 顾观文 武晔 孙守才 王顺吉 王英杰 林兴龙 吴昊彦 CAO Lai;GU Guanwen;WU Ye;SUN Shoucai;WANG Shunji;WANG Yingjie;LIN Xinglong;WU Haoyan(School of Earth Sciences,Institute of Disaster Prevention,Sanhe 065201,China;Hebei Provincial Key Laboratory of Earthquake Dynamics,Sanhe 065201,China;Langfang KeyLaboratory of Geodetection and Information Technology,Sanhe 065201,China)

机构地区：[1]防灾科技学院地球科学学院,河北三河065201 [2]河北省地震动力学重点实验室,河北三河065201 [3]廊坊市地球探测与信息技术重点实验室,河北三河065201

出　　处：《防灾科技学院学报》2025年第1期78-88,共11页Journal of Institute of Disaster Prevention

基　　金：河北省在读硕士研究生创新能力培养计划(CXZZSS2025146);中央高校创新团队项目(ZY20215136、ZY20215108);河北省高等学校科学技术研究项目(ZC2022056);廊坊市科技支撑计划项目(2023013184);国家自然科学基金(42174177)。

摘　　要：对主流神经网络在地震勘探速度建模中的应用现状、优势和不足进行分析,旨在为今后的高精度地震速度建模研究提供有益参考。其中,卷积神经网络(CNN)能够有效映射地震数据与速度之间的非线性关系;生成对抗神经网络(GAN)能扩展训练样本,增强泛化能力;递归神经网络(RNN)用于速度谱自动拾取;自编码神经网络(AE)能提高速度建模的抗噪能力;物理信息神经网络(PINN)融入了物理约束条件,可替代传统波动方程求解。每种方法各具特色,组合使用可优势互补,提升速度建模效率和精度。In this paper,aiming to provide the useful reference for future research on high-precision seismic velocity modeling,the current applications,advantages,and shortcomings of main neural network methods applied in seismic exploration velocity modeling are analyzed.Among these methods,Convolutional Neural Networks(CNN)can effectively map the nonlinear relationship between seismic data and velocity;Generative Adversarial Networks(GAN)can expand training samples and enhance generalization ability;Recurrent Neural Networks(RNN)are used for automatic picking of velocity spectra;Autoencoders(AE)can improve the noise resistance of velocity modeling;Physics-Informed Neural Networks(PINN)incorporate physical constraints,serving as alternatives to traditional wave equation solutions.Each neural network method has its own features.When combined these methods can complement each other's strengths and improve the efficiency and accuracy of velocity modeling.

关 键 词：深度学习 速度建模 地震勘探 神经网络 模型优化 

分 类 号：P315.3[天文地球—地震学]