Joint Learning for Seismic Inversion: An Acoustic Impedance Estimation Case Study (2006.15474v1)

Abstract: Seismic inversion helps geophysicists build accurate reservoir models for exploration and production purposes. Deep learning-based seismic inversion works by training a neural network to learn a mapping from seismic data to rock properties using well log data as the labels. However, well logs are often very limited in number due to the high cost of drilling wells. Machine learning models can suffer overfitting and poor generalization if trained on limited data. In such cases, well log data from other surveys can provide much needed useful information for better generalization. We propose a learning scheme where we simultaneously train two network architectures, each on a different dataset. By placing a soft constraint on the weight similarity between the two networks, we make them learn from each other where useful for better generalization performance on their respective datasets. Using less than 3$\%$ of the available training data, we were able to achieve an average $r^{2}$ coefficient of 0.8399 on the acoustic impedance pseudologs of the SEAM dataset via joint learning with the Marmousi dataset.

• The paper introduces a joint deep learning framework for acoustic impedance estimation in seismic inversion to address the challenge of limited well log data.
• The method uses two concurrently trained networks with soft weight sharing, temporal convolutional networks, and multi-task learning to enhance generalization.
• The approach achieves an average correlation of 0.8399 predicting acoustic impedance using less than 3% of training data, demonstrating high accuracy with limited resources.

Joint Learning for Seismic Inversion: An Acoustic Impedance Estimation Case Study

In the field of geophysics, seismic inversion plays a crucial role in constructing accurate subsurface models to aid in hydrocarbon exploration and production. The paper by Ahmad Mustafa and Ghassan AlRegib presents a methodological advancement in this domain, leveraging joint learning through a deep learning framework to enhance the accuracy and efficiency of acoustic impedance estimation. This method is particularly pivotal given the limited availability of well log data, a common issue in this field due to the high costs associated with drilling.

The traditional approach to seismic inversion relies heavily on large quantities of well-labeled data to establish a precise mapping from seismic data to rock properties. The limitations in well log availability pose a significant challenge, often leading to overfitting and suboptimal generalization in machine learning models. This paper introduces an innovative joint learning approach, whereby two identical network architectures are trained on different datasets concurrently. By imposing a soft constraint on the similarity of their weights, the networks learn mutually beneficial features. This setup facilitates improved generalization across datasets, even when only a small subset of available training data is utilized.

In a practical application of their method, the authors achieve an average correlation coefficient of 0.8399 in predicting acoustic impedance at non-well positions using less than 3% of the available training data. This statistical outcome underscores the efficacy of their approach in capturing key subsurface features, despite the limited training dataset size.

The methodological contribution of this paper revolves around a transfer learning scheme that effectively balances knowledge sharing and dataset-specific learning. By using temporal convolutional networks with multi-task learning (via simultaneous regression and reconstruction tasks), the networks can learn robust feature representations that account for both spatial context and temporal dependencies. The paper further demonstrates the scalability potential of this framework, as it is not confined to a specific number of training datasets, allowing for broad applicability and extension to other geological properties beyond acoustic impedance.

From a theoretical perspective, the introduction of soft weight sharing offers a flexible and effective mechanism for fostering generalization in neural network architectures, particularly within the context of seismic data interpretation. This paper positions itself as an important milestone towards overcoming data scarcity in seismic inversion, proposing a network training paradigm that capitalizes on related datasets to compensate for individual dataset limitations. Future developments may explore the utilization of this framework in other domains where labeled data acquisition is similarly challenging and costly, and its incorporation with more varied network architectures to further enhance generalization capabilities.