影响因子：

4.9

DOI码：

10.1016/j.cageo.2026.106277.

发表刊物：

Computers and Geosciences

摘要：

Magnetotellurics (MT) inversion is essential for exploring subsurface electrical resistivity structures, but remains challenging due to nonlinearity and ill-posedness. Gradient-descent methods are often sensitive to local minima and initial models, whereas nonlinear methods are computationally expensive. Motivated by recent advances in deep learning, we propose a physics-constrained implicit neural representation (INR) framework for 2D MT inversion. Unlike data-driven neural networks, the proposed method constrains the reconstruction through a data-misfit loss evaluated using a Maxwell-equation-based MT forward operator, rather than requiring labelled training data. The continuous and compositional structure of the neural network provides an implicit regularization effect, enabling a smooth and continuous representation while retaining flexibility for complex variations. Block-model experiments confirm the robustness of the method across the tested noise levels, network sizes, and initializations. In the synthetic layered model, INR achieves an acceptable data fit comparable to MARE2DEM and yields improved model-domain metrics, reducing root mean square error (RMSE) from 0.3338 to 0.2445 and increasing structural similarity index measure (SSIM) from 0.3282 to 0.6378. Application to the COPROD2 field data further confirms the capability of the proposed 2D INR-MT inversion to recover complex subsurface structures. These results highlight INR as a promising foundation for next-generation MT inversion and motivate future extensions to 3D geophysical imaging, particularly for applications such as geological carbon dioxide storage monitoring and high-resolution characterization of complex subsurface structures. Keywords: Magnetotelluric inversion; Deep learning; Implicit neural representations; Neural fields

合写作者：

Xavier Garcia, Eric Attias, Boyao Zhang, Fansheng Xiong, Jianxin Liu

第一作者：

Yanyi Wang

论文类型：

期刊论文

通讯作者：

Zhenwei Guo

卷号：

218

页面范围：

106277

ISSN号：

0098-3004

是否译文：

否

发表时间：

2026-09-27

收录刊物：

SCI