Semi-supervised Learning for Acoustic Impedance Inversion (1905.13412v1)

Abstract: Recent applications of deep learning in the seismic domain have shown great potential in different areas such as inversion and interpretation. Deep learning algorithms, in general, require tremendous amounts of labeled data to train properly. To overcome this issue, we propose a semi-supervised framework for acoustic impedance inversion based on convolutional and recurrent neural networks. Specifically, seismic traces and acoustic impedance traces are modeled as time series. Then, a neural-network-based inversion model comprising convolutional and recurrent neural layers is used to invert seismic data for acoustic impedance. The proposed workflow uses well log data to guide the inversion. In addition, it utilizes a learned seismic forward model to regularize the training and to serve as a geophysical constraint for the inversion. The proposed workflow achieves an average correlation of 98% between the estimated and target elastic impedance using 20 AI traces for training.

• The paper introduces a semi-supervised framework that integrates convolutional and recurrent neural networks for acoustic impedance inversion to reduce labeled data requirements.
• The method achieves an impressive 98% correlation between estimated and true elastic impedance using only 20 AI training traces on the Marmousi 2 model.
• The approach incorporates a learned forward seismic model to regularize training and enforce geophysical constraints, promising cost-effective and scalable inversion.

Semi-supervised Learning for Acoustic Impedance Inversion

The paper "Semi-supervised Learning for Acoustic Impedance Inversion" presents an exploration into a novel application of semi-supervised learning methodologies within the field of geophysics, specifically addressing the challenge of acoustic impedance (AI) inversion. Authored by Motaz Alfarraj and Ghassan AlRegib, this work aims to mitigate the significant data requirements typically necessary in deep learning models by employing a semi-supervised framework that integrates convolutional and recurrent neural networks.

Background and Methodology

Seismic inversion, the process of extracting rock properties from seismic reflection data, traditionally contends with challenges due to the heterogeneity and non-linearity of subsurface properties, resulting in ill-posed inverse problems. Classical approaches often depend on Bayesian inference to incorporate prior knowledge and context. However, the authors of this paper explore the application of machine learning methods, leveraging their ability to model complex non-linear mappings through neural architectures.

In this paper, the authors propose a semi-supervised learning framework underpinned by convolutional and recurrent neural networks to address seismic inversion tasks. The core advantage of this approach lies in its reduced dependency on labeled data, which represents a significant departure from supervised learning techniques that typically demand extensive labeled datasets for model training. The proposed model employs seismic and acoustic impedance traces as temporal data series, achieving inversion using neural network models guided by well-log data. Additionally, the model incorporates a learned forward seismic model to regularize training, thus imparting geophysical constraints crucial for accurate inversion.

Empirical Evaluation and Results

The empirical validation of this proposed method was carried out using the Marmousi 2 model, an extensively utilized testing ground in geophysics for inversion and imaging analysis. Results from applying the proposed method point to an impressive average correlation of 98% between the estimated and true elastic impedance using merely 20 training AI traces. The strong numerical results indicate the effectiveness of the approach, even when limited labeled data is available. A scatter plot analysis between predicted and actual AI also reveals linear correlations, indicating the robustness of the model’s predictions.

Implications and Future Directions

These findings have several implications. Practically, the semi-supervised framework opens up the possibility for more extensive and reliable seismic inversions with limited well-log data, thus reducing costs and improving accessibility in resource-limited settings. Theoretically, it provides a renewed understanding of how machine learning can be intersected with classical geophysical techniques to address inverse problems.

Given the strong correlations observed and the efficiency of the workflow, future developments could explore the extension of this methodology to full elastic inversion and beyond elastic impedance estimation. Such extensions could contribute substantially to the field of reservoir characterization and management, potentially impacting both exploratory and production phases in the oil and gas industries. Additionally, investigating the model's scalability and adaptability across different seismic datasets could further enhance its utility and deployment in various geological scenarios.

The proposed semi-supervised learning approach heralds a significant step towards more efficient and less data-hungry inversion techniques in geophysics, positioning it as a noteworthy development in the application of deep learning within the field of earth sciences.