Method for predicting subsurface features from seismic using deep learning dimensionality reduction for regression

Abstract

A method for training a backpropagation-enabled regression process is used for predicting values of an attribute of subsurface data. A multi-dimensional seismic data set with an input dimension of at least two is inputted into a backpropagation-enabled process. A predicted value of the attribute has a prediction dimension of at least 1 and is at least 1 dimension less than the input dimension.

Claims

1. A method for training a backpropagation-enabled regression process for predicting values of an attribute of subsurface data, the method comprising the steps of: (a) inputting a multi-dimensional seismic data set with an input dimension, n, of at least two into a backpropagation-enabled process; (b) convolving the multi-dimensional seismic data set using convolutional filters learned from the backpropagation-enabled process to produce a set of input feature maps having a convolutional dimension equal to the input dimension; (c) downscaling the set of input feature maps to produce a downscaled array of modified feature maps using the convolutional filters learned from the backpropagation-enabled process, so that at least one, but no more than n−1, dimension of the downscaled array is one; (d) upscaling the downscaled array to produce an upscaled array of further modified feature maps using the convolutional filters learned from the backpropagation-enabled process, wherein the upscaled array has the same dimension of the downscaled array, and (e) computing a predicted value of the attribute from the upscaled array, wherein the predicted value has a prediction dimension of at least 1 and is at least 1 dimension less than the input dimension.

2. The method of claim 1, wherein the input dimension is at least 2, and the prediction dimension is 1.

3. The method of claim 1, wherein the input dimension is at least 3, and the prediction dimension is selected from the group consisting of 1 and 2 dimensions.

4. The method of claim 1, wherein the input dimension is at least 4, and the prediction dimension is selected from the group consisting of 1, 2 and 3 dimensions.

5. The method of claim 1, wherein the input dimension is at least 5, and the prediction dimension is selected from the group consisting of 1, 2, 3 and 4 dimensions.

6. The method of claim 1, wherein the multi-dimensional seismic data is selected from the group consisting of 2D seismic data and 2D data extracted from seismic data of 3 or more dimensions.

7. The method of claim 1, wherein the multi-dimensional seismic data set is selected from the group consisting of 3D seismic data and 3D data extracted from seismic data of 4 or more dimensions.

8. The method of claim 1, wherein the multi-dimensional seismic data set is selected from the group consisting of 4D seismic data and 4D data extracted from seismic data of 5 or more dimensions.

9. The method of claim 1, wherein the multi-dimensional seismic data set comprises multiple attributes.

10. The method of claim 9, wherein the multiple attributes comprise 3 color channels.

11. The method of claim 1, wherein the predicted value is made on a 1D seismic array from an input dimension of at least of 2.

12. The method of claim 1, wherein the predicted value is made on a 2D seismic array from an input dimension of at least 3.

13. The method of claim 1, wherein the predicted value is made on a 3D seismic array from an input dimension of at least 4.

14. The method of claim 1, wherein the predicted value is made on a 4D seismic array from an input dimension of at least 5.

15. The method of claim 1, wherein the predicted value is made on a 5D seismic array from an input dimension of at least 6.

16. The method of claim 1, wherein the backpropagation-enabled process is a deep learning process.

17. The method of claim 1, wherein the backpropagation-enabled process is a supervised regression process, comprising the step of comparing attributes computed in a conventionally computed technique with the ones from a supervised regression technique.

18. The method of claim 1, wherein the backpropagation-enabled process is selected from the group consisting of supervised, semi-supervised, unsupervised processes and combinations thereof.

19. The method of claim 1, wherein the multidimensional seismic data set is comprised of seismic data selected from the group consisting of field-acquired data, synthetic data, and combinations thereof.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

(1) The method of the present invention will be better understood by referring to the following detailed description of preferred embodiments and the drawings referenced therein, in which:

(2) FIG. 1 is a representation of a Prior Art 2D Deep Learning (“DL”) input/output pair;

(3) FIG. 2 is a representation of a Prior Art 3D DL input/output pair;

(4) FIG. 3 is a representation of one embodiment of the present invention having a 3D seismic data input and a 2D prediction; and

(5) FIG. 4 illustrates another embodiment of the present invention.

DETAILED DESCRIPTION OF THE INVENTION

(6) The present invention provides a method for training a backpropagation-enabled regression process for predicting values of an attribute of subsurface data. Examples of attributes of subsurface data include any quantity derived from the input data such as, without limitation spectral content, energy associated with changes in frequency bands, signals associated with filters, including without limitation, noise-free filters, low-pass filters, high-pass filters, and band-pass filters, acoustic impedance, reflectivity, semblance, loop-based properties, envelope, phase, dip, azimuth, curvature and the like.

(7) Analysis of subsurface data, including seismic data, is important for improving efficiency and accuracy of hydrocarbon exploration. However, seismic data is often voluminous and subject to human error in interpretation. For example, in the interpretation of field-acquired data, faults are annotated by drawing a series of straight lines. But the fault may not fall exactly along the straight-line segments or interpreters may inadvertently interpolate segments into a single fault where multiple faults are present. Also, the presence of noise from, for example, seismic acquisition, seismic processing, an imaging process and, often, from combinations thereof, can adversely affect the ability to interpret the prospectivity of the region targeted by the subsurface data collection. In practice, this limitation is overcome by computing attributes from the subsurface data. Accordingly, it would be advantageous to leverage backpropagation-enabled processes for predicting values of an attribute of subsurface data.

(8) Examples of backpropagation-enabled processes include, without limitation, artificial intelligence, machine learning, and deep learning. It will be understood by those skilled in the art that advances in backpropagation-enabled processes continue rapidly. The method of the present invention is expected to be applicable to those advances even if under a different name. Accordingly, the method of the present invention is applicable to the further advances in backpropagation-enabled processes, even if not expressly named herein.

(9) A preferred embodiment of a backpropagation-enabled process is a deep learning process, including, but not limited to a convolutional neural network.

(10) In accordance with the present invention, the method for training the backpropagation-enabled process involves inputting a multi-dimensional seismic data set with an input dimension of at least two into the backpropagation-enabled process. The input multi-dimensional seismic data set may be 2D, 3D, 4D, 5D, and the like. The input 2D data set may, for example, be 2D seismic data or 2D data extracted from seismic data of 3 or more dimensions. Likewise, the input 3D data set may, for example, be 3D seismic data or 3D data extracted from seismic data of 4 or more dimensions. And, input 4D data set may, for example, be 4D seismic data or 4D data extracted from seismic data of 5 or more dimensions.

(11) The multidimensional seismic input data set may include field-acquired data, synthetic data, and combinations thereof.

(12) Once the data is inputted to the backpropagation-enabled process, a predicted value of the attribute is computed. In accordance with the present invention, the predicted value has a prediction dimension of at least 1 and is at least 1 dimension less than the input dimension. In the backpropagation regression process, the predicted value of the attribute is on the same grid as the input data or a subset thereof.

(13) In a preferred embodiment, the input dimension of the multi-dimensional seismic data set is at least 2 and the prediction dimension is 1. An example of a 1-dimensional predicted value is a ribbon, for example representing a line in a 2-dimensional slice or grid. The line may be, for example, a line in an x or y direction. Another example of a 1-dimensional predicted value is a trace, for example, a signal in the z-direction at a single specified xy-value.

(14) In another preferred embodiment, the input dimension of the multi-dimensional seismic data set is at least 3 and the prediction dimension is 1 or 2. In a further preferred embodiment, the input dimension is at least 4 and the prediction dimension is 1, 2 or 3. In yet another preferred embodiment, the input dimension is at least 5 and the prediction dimension is 1, 2, 3 or 4.

(15) Preferably, the predicted value is made on a 1D seismic array from an input dimension of at least 2. More preferably, the predicted value is made on a 2D seismic array from an input dimension of at least 3. In another embodiment, the predicted value is made on a 3D seismic array from an input dimension of at least 4. In yet another embodiment, the predicted value is made on a 4D seismic array from an input dimension of at least 5. Still another embodiment has a predicted value made on a 5D seismic array from an input dimension of at least 6. When the data is larger than 3D, it is generally 3D seismic data with attributes related to seismic survey acquisition or the result of multiple attribute computations. As an example, multiple attributes preferably comprise 3 color channels.

(16) Referring now to the drawings, FIG. 1 illustrates a Prior Art 2D Deep Learning (“DL”) input/output pair 1. The DL input/output pair 1 has input data 2 and predicted values 3 that are identically-sized 2D arrays to be used for training. The labels (not shown) in a supervised process are also 2D arrays with the same shape as the input data 2.

(17) Likewise, FIG. 2 illustrates a Prior Art 3D DL input/output pair 1 having input data 2 and predicted values 3 that are identically-sized 3D arrays to be used for training. The labels (not shown) in a supervised process are also 3D arrays with the same shape as the input data 2.

(18) FIG. 3 is a representation of one embodiment of a DL input/output pair for the method 10 of the present invention, wherein the input data 12 and predicted values 14 have different dimensions. Labels (not shown) in a supervised process have same dimensions and shape as predicted values 14. The predicted values 14 have at least one fewer dimension than the input seismic data 12 and are at least 1D.

(19) FIG. 4 is a representation of another embodiment of a training architecture 100 for a backpropagation-enabled process in the method of the present invention. In the example illustrated in FIG. 4, the open triangles represent 3D convolutions, while the solid triangles represent 2D convolutions. Meanwhile, the open arrows represent downscaling steps, while the solid arrows represent up scaling steps.

(20) A specific example is demonstrated in FIG. 4. However, it will be understood by those skilled in the art that the dimension of the input data, the size/shape of the input data, the number and size of convolutional filters used, the number of convolutions, the number of downscaling steps, the number of upscaling steps, and the number of feature maps created can be different. The multi-dimensional seismic data set is shown as 3D in the example of FIG. 4, but can be 4D, 5D, etc.

(21) For ease of discussion, the size of the input data set and filters for each step was selected to produce feature maps of the same size. However, this is not necessary in accordance with the method of the present invention. Each feature map dimension has a value whose maximum size is identical to the size of the corresponding dimension of its input data set and whose minimum size is the size of the corresponding input data set minus twice the integer value of the half-size of the corresponding dimension of the filter.

(22) As shown in the example of FIG. 4, a multidimensional seismic input data set 112 is 3D with a shape of 600×2000×5 pixels. The input data set 112 is convolved with 16 3D convolutional filters using weights that are learned via backpropagation-enabled deep learning. This creates 16 feature maps 116. In this case, each feature map 116 array has the same dimension as the input data set 112, in this case 600×2000×5. The set of 16 feature maps 116 is then convolved with 16 3D convolutional filters using weights that are learned via back-propagation enabled deep learning, resulting in a set of modified feature maps 118. This step modifies the content of the input feature maps 116 but does not change the dimension or shape.

(23) In a subsequent step, each modified feature map 118 is downscaled to produce a set of 32 feature maps 122. In this example, downscaling involves decimating the modified feature maps 118 in each dimension by outputting every other element to change the shape from 600×2000×5 to 300×1000×3. The set of 32 feature maps 122 is then convolved with 32 3D convolutional filters using weights that are learned via back-propagation enabled deep learning, resulting in another set of modified feature maps 124. This step modifies the content of the input feature maps 122 but does not change the dimension or shape.

(24) Next, the number of feature maps 124 is doubled from 32 to 64 feature maps 126 in a second downscaling step, where the shape of each feature map 124 is decimated by 2, i.e., from 300×1000×3 to 150×500×1. As noted above, the size and shape of the data may be different than shown in FIG. 4. However, in accordance with the present invention, the dimension, n, of the data should be downscaled by an appropriate number of downscaling steps so that at least one, but no more than n−1, of the array dimensions is one. In other words, the feature maps 126 are now 2D, although FIG. 4 shows the feature map with 3 dimensions and shape as 150×500×1 for clarity. When one of the dimensions of the set of feature maps is one element, the number of dimensions is reduced by one, such that the dimension is equally described as a 2D array with shape 150×500. The set of 64 feature maps 126 is then convolved with 64 2D convolutional filters using weights that are learned via back-propagation enabled deep learning, resulting in another set of modified feature maps 128. This step modifies the content of the input feature maps 126 but does not change the dimension or shape.

(25) Each feature map 128 is then upscaled to change the shape from 150×500 to 300×1000 in the set of feature maps 132 by infilling in each dimension plus additional elements between each element in the input 2D array of the set of feature maps 128. The set of 32 feature maps 132 is then convolved with 32 2D convolutional filters using weights that are learned via back-propagation enabled deep learning, resulting in another set of modified feature maps 134. This step modifies the content of the input feature maps 132 but does not change the dimension or shape.

(26) The predicted value 114 output has one (or more) fewer dimensions than the input data 112. As shown in FIG. 4, the shape of the remaining dimensions is the same as the input data—i.e., both are 600×2000. However, the shape of the predicted value 114 may also be smaller in the remaining dimensions, as long as the predicted value 114 is at least 1D. More specifically, a predicted value that is a single pixel or a single voxel is zero D and is outside the scope of the present invention.

(27) The backpropagation-enabled process may be supervised, semi-supervised, unsupervised or a combination thereof. In one embodiment, a supervised process is made semi-supervised by the addition of an unsupervised technique. As an example, the unsupervised technique may be an auto-encoder step.

(28) In a preferred embodiment, the backpropagation-enabled process is a supervised regression process. Preferably, the supervised regression process comprises the step of comparing attributes computed in a conventionally computed technique with the ones from a supervised regression technique.

(29) In another embodiment, the backpropagation-enabled process is an unsupervised regression process, comprising the steps of encoding and clustering the multi-dimensional seismic data.

(30) While preferred embodiments of the present invention have been described, it should be understood that various changes, adaptations and modifications can be made therein within the scope of the invention(s) as claimed below.