Full waveform inversion algorithm for subsurface physical property reconstruction using generative deep learning approach
The inverse problem in seismic exploration is addressed by leveraging both the Radon transform and deep learning techniques to reconstruct subsurface properties from seismic data. The method encompasses a workflow that employs Generative Adversarial Networks (GANs) alongside specialized neural network architectures called R-nets. The first R-net efficiently handles hyperbolic Radon Transform, aiding in data preprocessing, while the second R-net generates detailed subsurface models. These reconstructed properties include the velocity of pressure wave propagation, velocity of shear wave propagation, impedance, and density. This novel approach significantly enhances the accuracy and efficiency of the reconstruction process, particularly beneficial for pre-stack depth migration. The use of both Radon transform and generative deep learning greatly reduces the time and resources traditionally needed for depth-velocity model construction, leading to the generation of superior subsurface models faster.
1 . A method of full waveform seismic inversion, comprising:
storing, in a computer memory, geophysical data in the form of common midpoint (CMP)-sorted seismic time-series records obtained from a seismic survey of a subsurface region;
employing a two-stage U-Net-like convolutional neural network architecture, that includes neural networks R-net1 and R-net2;
training R-net1, implemented as a Generative Adversarial Network (GAN), to execute the hyperbolic Radon Transform on the CMP-sorted seismograms, using hyperbolic trajectories found in reflection events, thereby extracting RMS velocity data and reducing a dimensionality of the seismic time-series records;
training R-net2, implemented as a Generative Adversarial Network (GAN), that processes Radon images that contain RMS velocities of pressure waves in the subsurface region, into a subsurface parameter model; and
extracting, using a computer, a subsurface parameter model by processing the seismic time-series records using the trained R-net1 and R-net2.
2 . The method of claim 1 , wherein the geophysical data includes any of onshore seismic data, offshore seismic data, and transition zone seismic data.
3 . The method of claim 1 , wherein the extracted subsurface parameter model includes velocity of propagation of the pressure waves.
4 . The method of claim 1 , wherein the seismic time-series records are multi-component, and the extracted subsurface parameter model further comprises one or more of the following parameters: velocity of shear wave propagation, acoustic impedance, or the density of the subsurface region.
5 . The method of claim 1 , further comprising training the U-Net-like convolutional neural networks with synthetically generated subsurface physical models consistent with prior geological information and simulated geophysical data generated from synthetically generated subsurface parameter models.
6 . The method of claim 5 , further comprising generating the simulated geophysical data based on any of an acoustic wave equation, an elastic wave equation and coupled acoustic-elastic wave equations, and applying appropriate boundary conditions to the simulated geophysical data.
7 . The method of claim 5 , further comprising subjecting the simulated geophysical data to an augmentation process to generate an expanded set of data without additional modeling, wherein the augmentation process adds noise and implements gaps in signal registration.
8 . The method of claim 7 , wherein the training of the R-net1 and R-net2 neural networks includes utilizing a generator neural network that adopts a U-Net-like segmentation model.
9 . The method of claim 5 , wherein the training of the U-Net-like convolutional neural network uses separate convolutional neural networks as generator and discriminator components within a generative-adversarial approach (GAN).
10 . The method of claim 1 , further comprising supplementing the training of the U-Net-like convolutional neural network with an additional dataset specifically designed to incorporate prior geological knowledge about the subsurface region;
wherein the additional dataset includes elements well information, stratigraphy, subsurface structural patterns, and/or geophysical property ranges.
11 . The method of claim 1 , wherein the training of the R-net1 and R-net2 neural networks in the Generative Adversarial Network (GAN) approach is performed using a gradient descent algorithm or a stochastic gradient descent algorithm, to optimize the extraction of RMS velocity data from the CMP-sorted seismograms and the derivation of the velocity models and other subsurface physical property models.
12 . The method of claim 1 , further comprising using the subsurface parameter model for subsurface interpretation, hydrocarbon exploration, and/or hydrocarbon production process.
13 . The method of claim 12 , further comprising using the subsurface parameter model for a geophysical imaging process.
14 . The method of claim 12 , further comprising using the subsurface parameter model as a starting model for a geophysical inversion process.
15 . The method of claim 12 , wherein the derived subsurface pressure wave propagation velocity models are utilized as inputs for a pre-stack depth migration process, so as to provide imaging of the subsurface structures for geological interpretation and hydrocarbon exploration.
16 . The method of claim 12 , wherein the subsurface pressure wave propagation velocity models derived are employed in a time-to-depth conversion process, converting the time-based seismic data into a depth-based representation for geological interpretation.
17 . The method of claim 1 , further comprising providing manual control for the extraction process of the subsurface parameter model by allowing the user to set specific boundary conditions, including a range of model values, a complexity of the model, and a size of an output of the model.
18 . A system for full waveform seismic inversion, comprising:
a seismic station including sources and receivers that acquire geophysical data in the form of common midpoint (CMP)-sorted seismic time-series records; and a non-transitory computer readable storage medium, encoded with instructions, which, when executed by a processor, causes the processor to:
store, in a computer memory, geophysical data in the form of common midpoint (CMP)-sorted seismic time-series records obtained from a seismic survey of a subsurface region;
employ a two-stage U-Net-like convolutional neural network architecture, that includes neural networks R-net1 and R-net2;
train the R-net1, implemented as a Generative Adversarial Network (GAN), to execute the hyperbolic Radon Transform on the CMP-sorted seismograms, using hyperbolic trajectories found in reflection events, thereby extracting RMS velocity data and reducing a dimensionality of the seismic time-series records;
train the R-net2, implemented as a Generative Adversarial Network (GAN), that processes Radon images that contain RMS velocities of pressure waves in the subsurface region, into a subsurface parameter model; and
extract, using a computer, a subsurface parameter model by processing the seismic time-series records using the trained R-net1 and R-net2.