Deep learning electromagnetic inversion with convolutional neural networks

Abstract

Geophysical inversion attempts to estimate the distribution of physical\nproperties in the Earth's interior from observations collected at or above the\nsurface. Inverse problems are commonly posed as least-squares optimization\nproblems in high-dimensional parameter spaces. Existing approaches are largely\nbased on deterministic gradient-based methods, which are limited by\nnonlinearity and nonuniqueness of the inverse problem. Probabilistic inversion\nmethods, despite their great potential in uncertainty quantification, still\nremain a formidable computational task. In this paper, I explore the potential\nof deep learning methods for electromagnetic inversion. This approach does not\nrequire calculation of the gradient and provides results instantaneously. Deep\nneural networks based on fully convolutional architecture are trained on large\nsynthetic datasets obtained by full 3-D simulations. The performance of the\nmethod is demonstrated on models of strong practical relevance representing an\nonshore controlled source electromagnetic CO2 monitoring scenario. The\npre-trained networks can reliably estimate the position and lateral dimensions\nof the anomalies, as well as their resistivity properties. Several fully\nconvolutional network architectures are compared in terms of their accuracy,\ngeneralization, and cost of training. Examples with different survey geometry\nand noise levels confirm the feasibility of the deep learning inversion,\nopening the possibility to estimate the subsurface resistivity distribution in\nreal time.\n