Abstract Eastern margin of the Qinghai-Tibet plateau and western margin of the Yangtze craton has a complex crust-mantle structure and is an important region for studying the crust-mantle deformation mechanism. Because the Rayleigh surface wave group velocity has a strong constraint on the structure of the crust and upper mantle, a lot of researches have used the Rayleigh surface wave group velocity to invert the crust-mantle structure in this region, but the layered multi-parameter joint inversion and the uncertainty of the inversion result evaluation is still a research hotspot. Motived by Conditional Variational Autoencoder (CVAE) which is composed of a decoding and encoding process that uses a deep neural network to predict the variational distribution of parameters, we propose a crustal model by CVAE to attain the forward and inversion relationship between the group velocity of Rayleigh wave and the crustal model. In which the CVAE decoding corresponds to the forward process and the encoding corresponds to the inverse process. Moreover, the inversion results of this method are constrained from three aspects as follows: The network inputs and outputs have the same group velocities; The decoded outputs are normal distribution of hidden variables representing the crustal model; The mean values of the decoded output during CVAE training are the sampled crustal models. The proposed method can invert multi-parameters and evaluate uncertainty of the inversion efficiently. Based on the latest group velocity model, we invert for mean and variance of thickness in the upper, middle, and lower sedimentary cover, and of thickness, P-wave velocity, S-wave velocity and density in the upper, middle, and lower crystalline crust for eastern Qinghai-Tibet plateau and western Yangtze craton. Compared with the existing crustal models obtained by different methods, the results of this study are in good agreement. We conclude that CVAE-based group velocity inversion of the crustal model is a feasible and reliable method.

中文翻译：

---

基于条件变分自编码器的青藏高原东部和扬子克拉通西部地壳模型

摘要 青藏高原东缘和扬子克拉通西缘具有复杂的壳幔结构，是研究壳幔变形机制的重要区域。由于瑞利面波群速度对地壳和上地幔的结构有很强的约束，很多研究利用瑞利面波群速度反演该地区的壳幔结构，但层状多参数联合反演和反演结果评价的不确定性仍然是一个研究热点。由条件变分自动编码器 (CVAE) 驱动，它由解码和编码过程组成，使用深度神经网络来预测参数的变分分布，为了得到瑞利波群速度与地壳模型的正反关系，我们提出了CVAE的地壳模型。其中CVAE解码对应前向过程，编码对应逆向过程。此外，该方法的反演结果受到以下三个方面的约束：网络输入和输出具有相同的群速度；解码后的输出是代表地壳模型的隐藏变量的正态分布；CVAE 训练期间解码输出的平均值是采样的地壳模型。所提出的方法可以反演多参数并有效地评估反演的不确定性。根据最新的群速度模型，反演上、中、下层沉积盖层厚度的均值和方差，青藏高原东部和扬子克拉通西部上、中、下结晶地壳的厚度、纵波速度、横波速度和密度。与现有不同方法获得的地壳模型相比，本研究结果具有较好的一致性。我们得出结论，基于 CVAE 的地壳模型群速度反演是一种可行且可靠的方法。