Deep-learning inversion: a next generation seismic velocity-model building method (1902.06267v1)

Abstract: Seismic velocity is one of the most important parameters used in seismic exploration. Accurate velocity models are key prerequisites for reverse-time migration and other high-resolution seismic imaging techniques. Such velocity information has traditionally been derived by tomography or full-waveform inversion (FWI), which are time consuming and computationally expensive, and they rely heavily on human interaction and quality control. We investigate a novel method based on the supervised deep fully convolutional neural network (FCN) for velocity-model building (VMB) directly from raw seismograms. Unlike the conventional inversion method based on physical models, the supervised deep-learning methods are based on big-data training rather than prior-knowledge assumptions. During the training stage, the network establishes a nonlinear projection from the multi-shot seismic data to the corresponding velocity models. During the prediction stage, the trained network can be used to estimate the velocity models from the new input seismic data. One key characteristic of the deep-learning method is that it can automatically extract multi-layer useful features without the need for human-curated activities and initial velocity setup. The data-driven method usually requires more time during the training stage, and actual predictions take less time, with only seconds needed. Therefore, the computational time of geophysical inversions, including real-time inversions, can be dramatically reduced once a good generalized network is built. By using numerical experiments on synthetic models, the promising performances of our proposed method are shown in comparison with conventional FWI even when the input data are in more realistic scenarios. Discussions on the deep-learning methods, training dataset, lack of low frequencies, and advantages and disadvantages of the new method are also provided.

Overview of Deep-Learning Inversion for Seismic Velocity-Model Building

The paper in discussion presents a novel approach for seismic velocity-model building (VMB) leveraging the strengths of supervised deep fully convolutional neural networks (FCN). Traditional methods such as tomography and full-waveform inversion (FWI) have been pivotal in deriving seismic velocities but are hindered by computational demands, time consumption, and dependency on human-quality control. In contrast, this paper advances a ML methodology aimed at mitigating these issues through using deep learning (DL) to automate and accelerate the inversion process for seismic data.

Methodology

The core of the research lies in constructing a nonlinear mapping between raw multi-shot seismic data and their corresponding velocity models. The proposed DL method works in two main stages: training and prediction. During training, the network learns from synthetic seismograms paired with velocity models, establishing a robust mapping that eschews prior knowledge assumptions typically needed in conventional approaches. The network uses an FCN architecture, allowing for automatic extraction of complex features relevant to VMB without manual intervention or an initial velocity setup. Key modifications to the FCN, drawing from the UNet architecture, enable direct projection from the seismic data space to the velocity model space.

Numerical Experiments and Results

Extensive numerical tests show promising results. For testing, 2D synthetic models were used, generated both synthetically and drawn from Society of Exploration Geophysics (SEG) reference models. For the training dataset, the authors designed velocity models with a variety of layers and included arbitrary shapes such as salt bodies, striving for realistic scenarios to enhance network robustness. Comparisons with traditional FWI underscore the efficiency of the FCN-based method, notably reducing prediction times from minutes to mere seconds after training completion.

Moreover, the DL method proved somewhat tolerant to perturbations such as noise and alteration in seismic amplitude, maintaining output fidelity to ground truth values even with disruptions. However, predictions were sensitive to the presence of low-frequency information, a persistent challenge also faced by FWI.

Implications and Future Directions

From a practical perspective, the paper demonstrates a computationally efficient alternative to conventional VMB methods. The FCN architecture, with its rapid prediction capabilities post-training, positions itself as a formidable candidate for real-time seismic inversion applications. The findings encourage further expansion of the training dataset diversity and the exploration of more complex geological scenarios.

Theoretically, this work contributes to a growing body of research aiming to integrate DL techniques within geophysical applications. It challenges current inversion paradigms by proposing architectures that bypass many limitations of traditional approaches. Future research might delve into augmenting training datasets, potentially using generative adversarial networks for synthetic data generation. Additionally, leveraging transfer learning principles could facilitate application to different but related inversion tasks, broadening the utility and scope of DL within geoscience.

In conclusion, while substantial work remains to optimize and generalize this approach, this paper convincingly establishes a DL path for improving the agility and efficacy of seismic VMB, setting a firm foundation for progressive exploration in this domain.