Machine learning driven dispersion curve picking
A method for modeling a subterranean volume includes receiving seismic data comprising a signal, generating a semblance in the frequency-wavenumber domain for the seismic data, wherein the semblance represents a coherence of the signal in the frequency-wavenumber domain, extracting one or more wave energy modes in the semblance using a machine learning model trained to identify dispersion curves in the semblance based on a visible characteristic of the dispersion curves, and generating a model representing surface wave propagation based at least in part on the identified one or more wave energy modes.
1 . A method for modeling a subterranean volume, comprising:
receiving seismic data comprising a signal;
generating a semblance in a frequency-wavenumber domain for the seismic data, wherein the semblance represents a coherence of the signal in the frequency-wavenumber domain;
extracting one or more wave energy modes in the semblance using a machine learning model trained to identify dispersion curves in the semblance based on a visible characteristic of the dispersion curves, wherein extracting the one or more wave energy modes in the semblance using the machine learning model comprises:
using an encoder network of the machine learning model, the encoder network having initialized tensors that encode spatial features of an input semblance as strided convolution down-samples; and
using a decoder network to construct a segmentation mask as a filter at a same scale as the input semblance to be applied to the semblance in conjunction with identifying the dispersion curves in the semblance; and
generating a model representing surface wave propagation based at least in part on the identified one or more wave energy modes.
2 . The method of claim 1 , further comprising training the machine learning model using a training corpus of dispersion volumes and associated training semblances collected from a plurality of geographic regions including one or more near-surface conditions.
3 . The method of claim 2 , further comprising increasing a size of the training corpus by shifting one or more of the training semblances by a vertical distance, a horizontal distance, or both.
4 . The method of claim 2 , wherein training the machine learning model further comprises:
generating a binary mask for each of the training semblances, wherein the binary masks each include labels corresponding to one or more pixels of a training semblance associated with the respective binary mask, the one or more pixels each representing part of one or more wave energy modes; and
training the machine learning model based on pairs of the binary masks and semblances.
5 . The method of claim 4 , further comprising interpolating between the labels to extract a dispersion curve from spaced-apart labels.
6 . The method of claim 1 , further comprising training the machine learning model, wherein training the machine learning model comprises concatenating one or more of the spatial features of a layer of the encoder network to a corresponding layer of the decoder network using skip connections that skip one or more layers of the encoder network, the decoder network, or both.
7 . The method of claim 1 , wherein the one or more modes that are extracted include a plurality of modes, and wherein the method further comprises ordering the plurality of modes based on propagation speed.
8 . The method of claim 1 , further comprising attenuating surface wave noise based at least in part on the model.
9 . A non-transitory computer-readable medium storing instructions that, when executed by at least one processor of a computing system, cause the computing system to perform operations, the operations comprising:
receiving seismic data comprising a signal;
generating a semblance in a frequency-wavenumber domain for the seismic data, wherein the semblance represents a coherence of the signal in the frequency-wavenumber domain;
extracting one or more wave energy modes in the semblance using a machine learning model trained to identify dispersion curves in the semblance based on a visible characteristic of the dispersion curves, wherein extracting the one or more wave energy modes in the semblance using the machine learning model comprises:
using an encoder network of the machine learning model, the encoder network having initialized tensors that encode spatial features of an input semblance as strided convolution down-samples; and
using a decoder network to construct a segmentation mask as a filter at a same scale as the input semblance to be applied to the semblance in conjunction with identifying the dispersion curves in the semblance; and
generating a model representing surface wave propagation based at least in part on the identified one or more wave energy modes.
10 . The medium of claim 9 , further comprising training the machine learning model using a training corpus of dispersion volumes and associated training semblances collected from a plurality of geographic regions including one or more near-surface conditions.
11 . The medium of claim 10 , wherein the operations further comprise increasing a size of the training corpus by shifting one or more of the training semblances by a vertical distance, a horizontal distance, or both.
12 . The medium of claim 10 , wherein training the machine learning model further comprises:
generating a binary mask for each of the training semblances, wherein the binary masks each include labels corresponding to one or more pixels of a training semblance associated with the respective binary mask, the one or more pixels each representing part of one or more wave energy modes; and
training the machine learning model based on pairs of the binary masks and semblances.
13 . The medium of claim 12 , wherein the operations further comprise interpolating between the labels to extract a dispersion curve from spaced-apart labels.
14 . The medium of claim 9 , wherein the operations further comprise training the machine learning model, and wherein training the machine learning model comprises concatenating one or more of the spatial features of a layer of the encoder network to a corresponding layer of the decoder network using skip connections that skip one or more layers of the encoder network, the decoder network, or both.
15 . The medium of claim 9 , wherein the one or more modes that are extracted include a plurality of modes, and wherein the operations further comprise ordering the plurality of modes based on propagation speed.
16 . The medium of claim 9 , wherein the operations further comprise attenuating surface wave noise based at least in part on the model.
17 . A computing system, comprising:
one or more processors; and
a memory system comprising one or more non-transitory computer-readable media storing instructions that, when executed by at least one of the one or more processors, cause the computing system to perform operations, the operations comprising:
receiving seismic data comprising a signal;
generating a semblance in a frequency-wavenumber domain for the seismic data, wherein the semblance represents a coherence of the signal in a frequency-wavenumber domain;
extracting one or more wave energy modes in the semblance using a machine learning model trained to identify dispersion curves in the semblance based on a visible characteristic of the dispersion curves, wherein extracting the one or more wave energy modes in the semblance using the machine learning model comprises:
using an encoder network of the machine learning model, the encoder network having initialized tensors that encode spatial features of an input semblance as strided convolution down-samples; and
using a decoder network to construct a segmentation mask as a filter at a same scale as the input semblance to be applied to the semblance in conjunction with identifying the dispersion curves in the semblance; and
generating a model representing surface wave propagation based at least in part on the identified one or more wave energy modes.
18 . The system of claim 17 , wherein the operations further comprise:
training the machine learning model using a training corpus of dispersion volumes and associated training semblances collected from a plurality of geographic regions including one or more near-surface conditions, wherein training the machine learning model comprises:
generating a binary mask for each of the training semblances, wherein the binary masks each include labels corresponding to one or more pixels of a training semblance associated with the respective binary mask, the one or more pixels each representing part of one or more wave energy modes; and
training the machine learning model based on pairs of the binary masks and semblances; and
interpolating between the labels to extract a dispersion curve from spaced-apart labels.
19 . The method of claim 1 , further comprising:
consolidating the identified one or more wave energy modes into a single mode track;
ordering the identified one or more wave energy modes whereby the identified one or more wave energy modes in the single mode track are consistently numbered at associated analysis locations with a fundamental mode of the identified one or more wave energy modes being assigned a predetermined numerical identifier; and
performing a noise modelling step utilizing the ordering of the identified one or more wave energy modes as being directly mapped to an order of surface wave noise removal.