青藏高原东缘-扬子地台西缘地壳上地幔结构特征直接反映了汶川地震和芦山地震发生的深部地质原因，地震波层析成像和接收函数方法研究表明松潘—甘孜地块的地壳上地幔滑脱解耦，形成铲形逆冲推覆，是形成汶川地震和芦山地震的主要构造原因。这些地震方法在资料覆盖度高的地方分辨率高，在资料覆盖程度差或没有资料的地方采用外推方法得到，可信度相对较低。提出基于深度学习的混合密度神经网络方法用于地震波速度分布和间断面特征研究，不但给出了一个最佳拟合模型，而且还能对模型参数不确定性进行分析。搜集青藏高原东缘-扬子地台西缘地区最新数字地震台网波形数据，采用深度神经网络技术、求解基于贝叶斯方法的基阶和高阶面波频散曲线非线性反演，得到研究区地壳上地幔地壳厚度、S波速度结构、410km、660km不连续面的深度变化。进一步建立青藏高原东缘-扬子地台西缘地区高分辨率岩石圈三维模型，探讨产生汶川地震与芦山地震的深部动力学过程。

As the structure of crust and upper mantle in eastern Tibetan plateau-western Yangtz block reflect the result of Wenchuan and Lushan earthquakes, many literatures show that the reason is Songpan-Ganzi block decoupling from upper mantle and forming listric-thrust nappe structure based on result of seismic tomography and receiver function. These technologies are of high resolution with high data coverage, however, the results are attainted by extrapolating with low data coverage. Focusing on the problem of seismic tomography and discontinuity based on deep learning and mixed density neural network, we not only attain a best fit model also we can assess uncertainty of model parameters. Based on deep learn and mixed density neural network after collecting newest seismic wave form data in eastern Tibetan plateau-western Yangtz block, we take Bayesian inversion of fundamental mode surface data for moho depth and s-wave velocity and high mode for topography of 410km and 600km. Furthermore, based on these 3-D high resolution lithospheric models in eastern Tibetan plateau-western Yangtz block, we discuss deep geodynamics of Wenchuan and Lushan earthquakes.