GPU/CPU协同加速的双变网格最小二乘逆时偏移理论方法研究

项目名称： GPU/CPU协同加速的双变网格最小二乘逆时偏移理论方法研究

项目编号： No.41274124

项目类型： 面上项目

立项/批准年度： 2013

项目学科： 天文学、地球科学

项目作者： 黄建平

作者单位： 中国石油大学（华东）

项目金额： 70万元

中文摘要： 通过开展复杂构造下深部非常规油气藏的高精度最小二乘逆时偏移研究，建立一套适用于深部非常规油气藏成像的GPU/CPU协同加速的双变网格最小二乘逆时偏移的理论方法、优化算法、处理模块和适用技术，满足深部非常规复杂构造和岩性油藏地震成像的需求。逆时偏移无倾角限制、适应复杂速度场，将传统的逆时偏移纳入最小二乘反演框架，可进一步改进逆时偏移方法的成像噪音、深部成像振幅不均衡且深部成像分辨率较低等不足，研究内容主要包括：1、针对研究区域速度场分布特征的变网格速度剖分策略；2、形成基于多炮同时偏移的双变网格最小二乘逆时偏移理论方法；3、研发动态的相位编码算法，压制成像的串扰噪音；4、推导高效的预条件算子和去模糊化算子，降低反演迭代次数，加快收敛速度；5、基于GPU/CPU算法移植，使程序的加速比提高30倍，实现基于GPU/CPU协同加速的双变网格最小二乘逆时偏移理论方法及适用技术。

中文关键词： 最小二乘偏移；GPU/CPU协同；变网格；编码；混合编码

英文摘要： Through the research of high-precision least-squares reverse time migration (RTM) for the complex structure of unconventional oil and gas reservoirs, We want to establish a set of GPU/ CPU collaborative computing bivariate grid least squares RTM method, which is applicable to the needs of the seismic imaging of deep unconventional complex structure and lithology reservoir. RTM has no angle limit and adapts to any complex velocity model, by inducing the traditional RTM into the framework of the least-squares RTM can further improve the limitation of RTM which mainly contains the imaging noise, limited resolution and the amplitude is not balanced for deep reservoir imaging. In this project, we will mainly research the following issues: 1. grid triangulation strategy based on the velocity field changes; 2. the formation of the multi-source simultaneous least-squares migration; 3, R & D dynamic phase encoding algorithm to suppress the crosstalk noise of the imaging; 4. deriving an efficient pre-condition operator and defuzzification operator to reduce the number of inversion iterations to improve the convergence rate; 5. based on the GPU/ CPU algorithm transplant, the speedup ratio could be 30 times faster than the traditional CPU version which makes the algorithm be an appropriate technology for industrial usage. T

英文关键词： Least-squares migration；GPU/CPU joint method；variable grid size；phase encoding；hybrid phase encoding