Machine learning in seismic tomography

该项目的目的是根据机器学习的最新进展开发新的地震数据处理和成像方法，并将其应用于被动和主动源数据。就被动数据而言，剑桥大学自2006年以来一直在冰岛运行临时宽带地震阵列，并已汇集了一个世界级的数据集，其中包括与其独特构造背景有关的数万次地震的高质量记录（Ágústsdóttir等人，2016）。特定的目标区域包括最近活跃的火山下方的管道系统，例如在北方火山区发现的管道系统（格林菲尔德等人，2016年）。例如，阿斯克亚火山系统是冰岛最大的火山系统之一，最近一次喷发是在1961年，但我们直到最近才开始了解其内部结构和动态。这些火山地区还有尚未充分开发的地热能潜力，这是新开发的成像方法的另一个可能目标。此外，地震仪器高度集中在活动裂谷带，这意味着衰减层析成像等方法很适合对与地壳增生过程有关的结构进行成像。CGG的参与意味着学生还可以访问OBS和OBN数据，以测试新开发的方法和工作流程。学生要承担的一个有用的附带项目是从OBN阵列中反演地下结构的环境噪声表面波数据;这将是一个有趣的机器学习应用，可以与陆地上类似应用的最新结果进行比较。

The aim of the project is to develop novel seismic data processing and imaging methods based on recent advances in machine learning and apply them to passive and active source data. In the case of passive data, the University of Cambridge has been operating temporary broadband seismic arrays in Iceland since 2006, and has assembled a world-class dataset that includes high quality recordings of tens of thousands of earthquakes related to its unique tectonic setting (Ágústsdóttir et al., 2016). Particular target areas include the plumbing systems beneath recently active volcanoes, such as those found in the Northern Volcanic Zone (Greenfield et al. 2016). For example, the Askja volcanic system, which last experienced an eruption as recently as 1961, is one of the largest volcanic systems in Iceland, yet we have only recently begun to understand its internal structure and dynamics. These volcanic regions also have geothermal energy potential that has not yet been fully explored, which is another possible target for newly developed imaging methods. Moreover, the high concentration of seismic instruments over the active rift zone means that methods such as attenuation tomography are well placed to image structure related to crustal accretion processes. The involvement of CGG means that the student will also have access to OBS and OBN data on which to test the newly developed methodology and workflows. A useful side project for the student to undertake would be to invert ambient noise surface wave data from the OBN array for subsurface structure; this would be an interesting application of machine learning that could be compared to recent results from similar applications on land.