瞬变电磁法可用于地下电阻率等物性参数的高精度成像，在矿产、地下水勘探等领域已得到广泛应用。目前瞬变电磁数据成像基本是基于一维反演，且一般不考虑激电效应；传统地球物理反演方法得到的电阻率等参数在空间上分布也较为扩散，很难识别地质体类型。本项目拟采用基于广义等效介质理论的激电模型（GEMTIP）和非结构化矢量有限元法在时间域直接模拟含激电效应的瞬变电磁响应。采取子区域分解法将大型三维问题分解为多个小规模的子问题，实现基于子区域分解的高效并行；对采集时间道进行并行处理，进一步提高算法的效率。采用多进制（multinary）反演方法，在多进制反演中电阻率等参数只会在某些可能的值附近分布，获取的物性参数空间对比度较高，易于区分物性分界面位置。利用机器学习方法从基于空间约束的一维反演结果中提取三维多进制反演的参数及初始模型，构建新的瞬变电磁反演流程。预期瞬变电磁多进制反演可得到更接近实际的准地质模型。

Transient electromagnetic (TEM) method can be used to obtain the high resolution resistivity structures of the subsurface. This method has been widely used in mineral exploration, groundwater investigation. The conventional TEM imaging methods are mostly based on one dimensional (1D) inversion and the induced polarization (IP) effects are usually ignored. In this project, we adopt the finite element method based on unstructured tetrahedral mesh to simulate the TEM response, directly in time domain. We consider the GEMTIP (generalized effective medium theory of induced polarization) model for IP effects modeling. By using unstructured tetrahedral discretization, we can accurately simulate topography and other complex geological structures. We consider the sub-domain strategy and divide a large 3D problem into a series of affordable small 3D problems. The numerical computation of such small 3D problems will be parallelized in modern cluster. To further accelerate the computation, we will implement the parallelization over TEM time channels. Within the framework of this approach, the early gate and late gate TEM data will be divided into several groups and they will be simulated independently with different time step sizes. The resistivity structures obtained from conventional inversion method are usually diffusive and it is hard to identify the rock types from the inversion result. In order to solve this problem, we propose to implement the multinary inversion method. In multinary inversion, the recovered physical parameters, such as resistivity, can only be distributed nearby several discrete values. The physical parameters obtained from multinary inversion is sharper and it is easier to identify the interface between different geological units. Before the 3D multinary inversion, we will first implement the 1D spatial constrained inversion (SCI). Using the machine learning approach, we could obtain a quasi-3D initial model and multinary parameters for 3D multinary inversion. As a result, we propose a new inversion flow from 1D SCI inversion to 3D multinary inversion. Using this new inversion method, we expect to obtain a realistic quasi-geological model with clearly boundary information.