密集台阵和深度学习方法揭示的2016年门源M_(S)6.4地震发震构造特征  

Revealing the seismogenic structural characteristics of the Menyuan M_(S)6.4 earthquake in 2016 based on dense array and deep learning

作　　者：任超 孙安辉[1] 王伟涛[2] 于子叶 洪顺英[1] 李澍辰 REN Chao;SUN AnHui;WANG WeiTao;YU ZiYe;HONG ShunYing;LI ShuChen(Institute of Earthquake Forecasting,China Earthquake Administration,Beijing 100036,China;Institute of Geophysics,China Earthquake Administration,Beijing 100081,China)

机构地区：[1]中国地震局地震预测研究所,北京100036 [2]中国地震局地球物理研究所,北京100081

出　　处：《地球物理学报》2025年第4期1287-1303,共17页Chinese Journal of Geophysics

基　　金：国家自然科学基金项目(42474134,41974050,42374006),地震预测研究所基本科研业务费专项(CEAIEF2023030104,CEAIEF20230201)联合资助.

摘　　要：针对近些年多次发生破坏性强震的青海门源地区,采用密集的中国地震科学探测台阵二期项目(ChinArray Ⅱ)的波形数据,应用三种深度学习模型(BRNN、EQTransformer、PhaseNet)完成震相拾取工作,并利用震相关联与地震精定位方法,优化了区域一维速度模型,最终获得了2016年门源M_(S)6.4地震高精度的地震序列目录.综合大地测量和震源机制解分析结果,发震断层应为一条次级盲断裂——冷龙岭北侧断裂.余震分布显示断层面浅部较为直立,深部倾角约为40°,倾向西南,并在13~17 km深度与冷龙岭断裂带交汇.交汇区局部复杂的断层几何特征、介质属性以及能量释放等综合因素,阻碍了2022年门源M_(S)6.9地震沿冷龙岭断裂向东继续破裂.密集台阵和机器学习方法将该时期研究区域的最小完备震级,由台网目录的ML0.8降为ML0.6.主震前20天内无前震活动,推测2016年门源M_(S)6.4地震序列是主震—余震序列类型.本文评估了深度学习模型在研究区的适配性能:EQTransformer模型在震相拾取数上均衡性表现较好,可定位到更多的地震;BRNN模型得到的结果走时残差最小.Several destructive earthquakes occurred in Menyuan County,Qinghai province in recent years.We apply three deep learning models(BRNN,EQTransformer,PhaseNet)on dense waveform data from ChinArrayII to pick seismic phases.We obtain a high-precision catalog for the 2016 Menyuan M_(S)6.4 earthquake sequence based on the seismic phase association and earthquake precision relocation methods and optimize the regional 1-D initial velocity model.With comprehensive geodesy and focal mechanism analysis results,we infer that the seismogenic fault for this earthquake is a blind secondary fault,the northern Lenglongling fault(LLLF).The aftershock distribution infers that the seismogenic fault plane might be relatively upright in the shallow part of the fault while southwest-dipping with about 40°in the deep and intersecting with the LLLF at the depths of 13~17 km.We proposed that the 2022 Menyuan M_(S)6.9 earthquake rupture eastward further along the Lenglongling fault might be halted by fault geometry complexity,medium properties,and energy release of the intersection zones.Applying machine learning methods with a dense seismic array helps reduce the minimum magnitude of completeness from ML0.8 to ML0.6 for the study area.There was no foreshock activity in the twenty days before the mainshock,thus we speculate that the 2016 Menyuan M_(S)6.4 earthquake sequence was a mainshock-aftershock sequence type.This paper evaluates the adaptability performance of different deep learning models in the study area.The results show that the EQTransformer model has a better balance in the number of picking phases and can locate more earthquakes,while the BRNN model obtains the smallest travel time residual.

关 键 词：密集台阵 深度学习 地震重定位 最小一维速度模型 门源地区 断层几何特征 

分 类 号：P315[天文地球—地震学]