ML 02 Pre-Stack Seismic Inversion With Deep Learning: Y. Zheng, Q. Zhang BP, BP America

Uploaded by

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

61 views

ML 02 Pre-Stack Seismic Inversion With Deep Learning: Y. Zheng, Q. Zhang BP, BP America

Uploaded by

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 4

ML 02

Pre-Stack Seismic Inversion With Deep Learning

Y. Zheng1*, Q. Zhang2
1
BP, 2 BP America

Summary
We present a study of seismic inversion using deep learning tools. The purpose is to investigate the feasibility of
using neural networks to construct acoustic and elastic earth models directly from pre-stack seismic data. Training
and testing of the neural networks are done using thousands of synthetic 1D earth models and seismic gathers.
We use 2 different types of neural network architectures in our numerical experiments to investigate seismic
inversion in different geological scenarios. In both cases, the quality of the prediction is comparable with that
obtained from conventional model building processes as such as travel-time and waveform inversion methods.
The predicted earth models contain abundant low and medium wave-number information. In terms of
performance, training took only less than 30 minutes on 4 GPUs whilst prediction adds negligible cost.

First EAGE/PESGB Workshop on Machine Learning

29-30 November 2018, London, UK
Introduction

The inversion of raw seismic data is typically treated as an optimization problem, during which an
inversion algorithm attempts to find an earth model that minimizes the difference between observed
and modelled data. Methods such as full waveform inversion (FWI), whilst effective in many cases,
often suffer from convergence issues and incur heavy computational cost. Recently various machine
learning techniques, particularly deep neural network (DNN), are being investigated for solving the
seismic inversion problem (Lewis and Vigh 2017, Araya-Polo et al. 2018, and Richardson 2018).
Here, we provide a proof-of-concept study of using deep learning techniques to predict subsurface
parameters in both acoustic and elastic 1D media. The goal is to assess the feasibility of waveform
inversion cast as a machine learning problem, where a DNN is trained to learn the non-linear
relationship between recorded seismic data and earth models.

Method

The application of DNN to seismic inversion can be largely grouped into two different approaches.
One approach is to embed the concept of deep learning into a conventional inversion framework
where components of the workflow such as forward modelling and gradient computation are replaced
by deep learning algorithms (Richardson 2018).

A more specific and pragmatic approach is to replace the entire inversion process by a training
process, allowing a DNN to find directly a non-linear function that transforms the input seismic data
to a model. This function is ‘flexible’ in the sense that it evolves with training. Here we present a
workflow of this deep learning approach. As illustrated in Figure 1, we constructed a set of random
synthetic 1D earth models and their corresponding pre-stack seismic data. A large subset is used to
train the network and the remaining data are used to evaluate its performance. We show the feasibility
of this approach through two different synthetic experiments: 1) Training a convolutional neural
network (CNN) to predict P-wave velocity in layered media; and 2) Training a multiple layer
perceptron (MLP) to predict elastic profiles in constrained layered media.

Figure 1 Comparison of a general classical inversion (a) to a deep learning workflow (b).

DNN experiments

In the first experiment, we assess a CNN for predicting P-wave velocity in acoustic 1D media. The
training data consist of 10,000 velocity profiles (labels) with random layer thickness and velocity, and
synthetic shot gathers (features) are derived from acoustic FD modelling (3 to 10Hz bandwidth). Also
2000 samples were generated for validation and test. Shot gathers are fed into the CNN as 2D images.
Training is done in mini-batches of 1000 samples across 4 GPUs, lasting about 30 minutes in total. As
shown in Figure 2, the predicted profiles contain relatively accurate back ground velocity structure,
but performance generally degrades with depth. Overall, the vertical resolution is comparable to what
would be obtained from travel-time tomography. In addition to using shot gathers, we also trained our
CNN using semblance and find the results very similar.

First EAGE/PESGB Workshop on Machine Learning

29-30 November 2018, London, UK
Figure 2 Predicted P-wave velocity profiles for 8 randomly selected test samples by CNN.

In the second example, we investigate another neural network architecture in MLP and attempt to
predict 3 elastic parameters (Vp, Vs, and density) simultaneously. Training labels are synthesized by
applying Gaussian processes to a set of true well data in a layer-by-layer manner, resulting in a suite
of elastic models with different yet statistically consistent vertical profiles. Shot gathers are created
using the reflectivity method (Kennett, 1983) to simulate elastic response (Uz) with a bandwidth of
3Hz to 60Hz. Training data consist of 5000 samples, 500 of which are used for validation and testing.
The duration of training across 4 GPUs takes about 20 minutes and prediction of 250 test samples a
on single CPU takes less than 3 seconds. Figure 3 shows the results of 2 random test cases and a
comparison with logs at a well location. The predicted elastic parameters show good match with the
truth, as the background trends and the strong impedance contrasts are recovered. It is noted that the
neural network is somewhat insensitive to the added white noise in the models used for training, but
prediction is quite sensitive to the seismic wavelet.

Figure 3 Predicted elastic models for 2 random tests (a and b) and at well location (c) by MLP.

Conclusions

We demonstrate the concept of using deep for seismic waveform inversion in 1D layered media. Our
synthetic examples show that DNNs may be used to build velocity models, with vertical resolution
comparable with that obtained from travel-time tomography and presumably low-frequency FWI. It is
also evident that DNN offers computational efficiency that may be desirable for elastic inversion.
Further investigation is needed for applications to 2D and 3D field data.

Acknowledgements

We would like to thank BP for the permission to publish this work.

First EAGE/PESGB Workshop on Machine Learning

29-30 November 2018, London, UK
References

Lewis, W and Vigh, D., [2017] Deep learning prior models from seismic images for full-waveform
inversion. 87th Annual International Meeting, SEG, Expanded Abstract, 1512-1517

Araya-Polo, M. Jennings, J., Adler, A. and Dahlke, T., [2018] Deep-learning tomography. The
Leading Edge, 37, 58-66

Richardson, A., [2018] Seismic Full-waveform Inversion Using Deep Learning Tools and
Techniques. ArXiv e-prints 1801.07232

Kennet, B.L.N., [1983] Seismic wave propagation in stratified media. Cambridge University Press.