Deep learning electromagnetic inversion with convolutional neural networks (1812.10247v1)

Abstract: Geophysical inversion attempts to estimate the distribution of physical properties in the Earth's interior from observations collected at or above the surface. Inverse problems are commonly posed as least-squares optimization problems in high-dimensional parameter spaces. Existing approaches are largely based on deterministic gradient-based methods, which are limited by nonlinearity and nonuniqueness of the inverse problem. Probabilistic inversion methods, despite their great potential in uncertainty quantification, still remain a formidable computational task. In this paper, I explore the potential of deep learning methods for electromagnetic inversion. This approach does not require calculation of the gradient and provides results instantaneously. Deep neural networks based on fully convolutional architecture are trained on large synthetic datasets obtained by full 3-D simulations. The performance of the method is demonstrated on models of strong practical relevance representing an onshore controlled source electromagnetic CO2 monitoring scenario. The pre-trained networks can reliably estimate the position and lateral dimensions of the anomalies, as well as their resistivity properties. Several fully convolutional network architectures are compared in terms of their accuracy, generalization, and cost of training. Examples with different survey geometry and noise levels confirm the feasibility of the deep learning inversion, opening the possibility to estimate the subsurface resistivity distribution in real time.

• The paper introduces a fully convolutional neural network framework that quickly estimates subsurface anomaly properties from EM data, bypassing costly gradient computations.
• It employs extensive 3-D synthetic datasets and IoU metrics to rigorously compare FCN architectures, demonstrating robust performance even in noisy conditions.
• The research offers a scalable real-time inversion tool that enhances CO2 storage monitoring and informed decision-making in geophysical exploration.

Deep Learning Electromagnetic Inversion with Convolutional Neural Networks

The paper "Deep learning electromagnetic inversion with convolutional neural networks" by Vladimir Puzyrev presents a paper on the application of deep learning (DL) methods, specifically convolutional neural networks (CNNs), for electromagnetic (EM) inversion. Geophysical inversion, which estimates the distribution of physical properties within the Earth from surface observations, often grapples with the challenges of nonlinearity and nonuniqueness. This work explores CNNs as a means to overcome these limitations, applying them to the domain of EM inversion with promising results.

Overview of Methodology and Results

The paper details a novel approach using fully convolutional neural networks (FCNs) trained on large synthetic datasets generated through comprehensive 3-D simulations. These networks, once trained, are proficient in estimating the position, dimensions, and resistivity properties of subsurface anomalies, with a key application being the monitoring of $\text{CO}_\text{2}$ storage sites. Notably, this method circumvents the need for gradient computations, a significant advantage over traditional approaches which are computationally expensive and time-intensive.

Several FCN architectures are rigorously compared, evaluating their accuracy, generalization capabilities, and training costs. The analysis does not merely focus on theoretical performance but validates the applicability through realistic examples, demonstrating that the trained networks can effectively handle variations in survey geometry and noise levels. This evidences the practical feasibility of the proposed DL inversion framework, which can achieve real-time predictions, hence offering significant improvements in computational efficiency.

The networks achieved substantial predictive accuracy on test datasets, with Intersection over Union (IoU) metrics indicating robust model generalization even with noisy data. This transition from deterministic and probabilistic inversion methods to a data-driven DL approach represents a pivotal shift, aimed at reducing the computational demands and enhancing the reliability of EM inversion tasks.

Implications and Future Directions

From a practical standpoint, this research introduces a scalable and efficient tool for real-time subsurface monitoring, particularly advantageous in applications such as $\text{CO}_\text{2}$ storage monitoring. The rapid estimation of resistive properties allows for better-informed decision-making processes in geophysical exploration and environmental monitoring.

Theoretically, this work underscores the potential of DL, specifically CNNs, to resolve geophysical inverse problems traditionally deemed prohibitive due to their complexity. The flexibility of the FCN architecture demonstrated here paves the way for further advancements in geoscientific modeling tasks extending beyond the scope of EM inversion—hidden dependencies and complex mapping relationships characteristic of high-dimensional inverse problems can now be effectively unraveled through tailored DL architectures.

Future research could explore extending this approach to accommodate anisotropic models or incorporating transfer learning techniques to broaden the network's applicability across different geophysical paradigms. Additionally, reformulation of the loss functions and further experimentation with retrospective and adversarial frameworks could refine and enhance the robustness of the inversion results, making DL an indispensable tool in computational geosciences.

Thus, this paper holds significant promise for both the geophysics community seeking efficient computational methods and the broader DL field, showing how domain-specific neural architectures can transform problem spaces traditionally ruled by conventional algorithms.