Estimation of Acoustic Impedance from Seismic Data using Temporal Convolutional Network (1906.02684v1)

Abstract: In exploration seismology, seismic inversion refers to the process of inferring physical properties of the subsurface from seismic data. Knowledge of physical properties can prove helpful in identifying key structures in the subsurface for hydrocarbon exploration. In this work, we propose a workflow for predicting acoustic impedance (AI) from seismic data using a network architecture based on Temporal Convolutional Network by posing the problem as that of sequence modeling. The proposed workflow overcomes some of the problems that other network architectures usually face, like gradient vanishing in Recurrent Neural Networks, or overfitting in Convolutional Neural Networks. The proposed workflow was used to predict AI on Marmousi 2 dataset with an average $r^{2}$ coefficient of $91\%$ on a hold-out validation set.

• The paper introduces a Temporal Convolutional Network (TCN) methodology for estimating acoustic impedance from seismic data, addressing limitations found in previous RNN and CNN approaches like gradient vanishing and overfitting.
• The TCN leverages dilated convolutions and temporal blocks, demonstrating efficacy by training on a small dataset (19 traces) from Marmousi 2 with supervised learning using MSE.
• Results show high predictive accuracy (91% average r2 on validation) and robust generalization, presenting a scalable framework for predicting subsurface properties from seismic data.

Estimation of Acoustic Impedance from Seismic Data using Temporal Convolutional Network

The paper "Estimation of Acoustic Impedance from Seismic Data using Temporal Convolutional Network" by Mustafa, Alfarraj, and AlRegib presents a methodology for acoustic impedance (AI) estimation utilizing Temporal Convolutional Networks (TCNs), addressing key challenges faced by earlier approaches that leverage Recurrent Neural Networks (RNNs) and Convolutional Neural Networks (CNNs). AI is a vital property in exploration seismology as it allows for the identification of subsurface structures, which are critical for hydrocarbon exploration. Traditional AI estimation techniques often struggle with the nonlinear mapping from seismic data to rock properties. This paper explores a sequence modeling methodology where TCNs are employed to predict AI from seismic data by overcoming issues such as gradient vanishing observed in RNNs, and overfitting prevalent in CNNs when applied to limited training datasets.

The research builds on previous efforts integrating supervised machine learning algorithms, which typically involve training on paired seismic and well-log datasets, to extrapolate physical properties across larger seismic volumes. However, these approaches are often limited by the amount of available training data and the complexity of feature mapping required to capture long-term dependencies and local variations in seismic data.

TCNs, characterized by dilated convolutions and temporal blocks, allow for efficient sequence modeling with a broader receptive field, addressing the shortcomings of RNNs and CNNs. The implementation leverages temporal block architecture with layers of convolution, dropout, weight normalization, and skip connections to achieve robust feature representation without an exponential increase in learnable parameters. By employing dilated convolutions, the network successfully maintains a large receptive field necessary for capturing extensive sequential dependencies.

Methodologically, the research employed TCNs on the Marmousi 2 dataset with encouraging outcomes. The proposed network architecture incorporates a series of six temporal blocks, optimizing a blend of kernel sizes and layers to balance high and low-frequency information capture. Within this framework, seismic data is processed as sequential input, and the network maps seismic to AI traces through a supervised learning regimen, efficiently training on a mere 19 traces of the dataset using Mean Square Error (MSE) as the loss function.

Remarkably, the trained TCN exhibited significant accuracy with an average $r^2$ coefficient of 91% on a validation set, showcasing its predictive capability. Pearson's Correlation Coefficient also reported similarly high values both in training and validation datasets, suggesting the model's robust generalization capability beyond its initial training data. The resulting AI predictions demonstrated high visual fidelity with the actual dataset and aligned well with ground truth values—reaffirming the model’s efficacy in structural delineation from seismic inputs.

The implications of this research extend beyond AI estimation, potentially offering a scalable and adaptable framework for predicting other subsurface properties, enhancing the utility of machine learning in geophysical explorations. The presented TCN-based model sets a precedent for future exploration in the development of computationally efficient, accurate seismic data inversion processes. Future work could involve further exploring the adaptability of TCNs to complex geophysical datasets, examining their applicability to other petrophysical properties, and potentially adapting similar methodologies to include real-time data processing capabilities.

In conclusion, this paper contributes a substantial advancement in applying temporal sequence modeling to geophysical data interpretation, showing promise for broad applications in exploration seismology and similar domains.