Research paperCrosstalk-free simultaneous-source full waveform inversion with normalized seismic data
Introduction
Full waveform inversion (FWI) constructs the medium parameters (Tarantola, 1984) by minimizing the misfit between the observed and simulated data. Its high resolution has attracted a great attention (Mora, 1987, Plessix, 2006, Guitton and Díaz, 2012, Operto et al., 2013, Xu and McMechan, 2014, Chen et al., 2016) since it was proposed. However, compared to the conventional ray-based travel-time tomography, the extremely low efficiency of FWI caused by the great computation amount is still a challenge for its wide application. The parallel algorithms (Brossier, 2011, Kim et al., 2013, Liu et al., 2013, Shin et al., 2014, Yang et al., 2014, Zhang et al., 2014, Liu et al., 2015, Fabien-Ouellet et al., 2017, Gui et al., 2017, Li et al., 2017, Xu et al., 2018, Wu et al., 2019, Gu et al., 2019, Wang et al., 2019, Fang et al., 2020) are usually used to improve its efficiency, but these algorithms are based on the high-performance-computing ability of computer devices and do not essentially reduce the computation amount.
To reduce the amount of computation, the simultaneous-source algorithms (Romero et al., 1999, Bansal et al., 2013) are the crucial strategies. Unfortunately, the conventional simultaneous-source method has two drawbacks: one is the crosstalk artifacts, which severely contaminates the inversion results; the other is the fixed acquisition assumption (Krebs et al., 2009) for all sources in a super shot, which is inapplicable to mobile streamer seismic data.
In order to overcome the false fixed-spread assumption, the global correlation-norm objective function (Choi and Alkhalifah, 2012) was proposed for simultaneous-source FWI. But Son et al. (2014) pointed out that the correlation operation (Choi and Alkhalifah, 2012) might distort the seismic data and developed an alternative approach by modifying the limited offset of observed data to the full offset. Nevertheless, when tens or hundreds of sources are encoded simultaneously, this alternative algorithm would be complicated and difficult to realize with little efficiency improvement.
As for the problem of crosstalk, Krebs et al. (2009) proposed to change the code sequence within a super shot at each iteration to attenuate the artifacts. Xue et al. (2017) and Zhang et al. (2018a) applied a shaping regularization constraint to the model parameter gradient to reduce the crosstalk influence, but the smoothing process would decrease the resolution of inversion results. Wu and Bai (2018) adopted a low-rank constraint to improve the resolution of simultaneous-source migration. Overall, the aforementioned strategies cannot eliminate the crosstalk artifacts essentially.
FWI can be considered as a data waveform matching process (Plessix, 2006) by iterative optimization algorithm. Therefore, the wavelet is crucial for a successful inversion because wrong wavelet would induce the waveform mismatching problem even the velocity is accurate (Zhang et al., 2016). However, the common-used direct arrival is not the true wavelet (Song et al., 1995) and the iterative source estimation (Xu et al., 2006) for each source within a super shot would increase the additional computation amount to counteract the efficiency of simultaneous-source method (Choi and Alkhalifah, 2011). In order to remove the inversion or estimation of source wavelet, two main strategies (Zhang et al., 2016) have been proposed, that is, the deconvolution (Zhou and Greenhalgh, 2003) and convolution (Choi and Alkhalifah, 2011, Zhang et al., 2016) methods. Both strategies need to choose a reference to eliminate or unify the source wavelet information. When using the conventional simultaneous-source algorithm and each source wavelet within a super shot having different signal functions, it is difficult to accomplish the source-independent FWI because of the simultaneous-source crosstalk and the acquisition mismatch between the simulated and encoded observed simultaneous-source data.
In the following context, we provide a novel crosstalk-free and source-independent simultaneous-source inversion algorithm with the normalized seismic data by deconvolution. The proposed algorithm uses the harmonic sine functions (Nihei and Li, 2007) as the source wavelet signals and encoding operators to conduct the time-domain simultaneous-source wavefield simulation, by which we can deblend the simultaneous-source wavefield. Therefore, the proposed algorithm can be naturally suitable for marine mobile streamer data. Likewise, it is easy to eliminate or unify the wavelet information for every source data without simultaneous-source crosstalk. The adjoint wavefield calculated by backward propagating the simultaneous-source residuals are also deblended so that the gradient can be calculated without the contamination of crosstalk artifacts. The whole deblending processes take only little time of the inversion.
In the following context, we first introduce the generation of crosstalk artifacts and make a detailed description for the proposed simultaneous-source algorithm to avoid this problem. Then we implement the source-independent simultaneous-source inversion with the normalized data by deconvolution. Next, we provide the deblending strategy to ensure the accuracy and adopt the Graphics Processing Unit (GPU) device to further enhance the efficiency of our algorithm. Finally, we use the synthetic examples to perform the feasibility analysis to our algorithm, and then give the corresponding discussion and conclusions.
Section snippets
Simultaneous-source crosstalk artifacts
The object function of FWI is built based on the following equation, where the superscripts of and represent the observed and simulated data, means the medium parameter, denotes the wavelet signal of the th source, is the total source number.
In order to enhance the inversion efficiency, the simultaneous-source algorithm is one crucial strategy. The corresponding objective function (Krebs et al., 2009) is expressed as
Examples
The Overthrust model is used to perform the simultaneous-source inversion tests. One hundred sources are set evenly at the surface. The temporal sampling interval is 1.0 ms and the length of recording seismogram is 8.0 s. The true velocity shown in Fig. 2(a) and the Ricker wavelet with 15.0 Hz dominant frequency are used to generate the unblended observed data. Fig. 2(b) shows the initial velocity for FWI, which is generated by the true velocity (Fig. 2(a)) after smoothing with a square spatial
Discussions
As for the frequency selection within a super shot for each source wavelet, one can choose as Eqs. (15)–(17), or by padding to double simulation time. It should be noted that different-frequency harmonic wavelets composes a band-limited super wavelet, so the number of encoded sources are restricted to assure the merit of multi-scale inversion scheme. In this paper, we do not perform analysis to quantitatively choose the optimal number of encoding sources. A suggestion is that one can
Conclusions
We provide a novel simultaneous-source algorithm with normalized seismic data for FWI. Sine harmonic functions are used as the wavelets and encoding operators. Based on the trigonometric orthogonality, the blended wavefield are deblended so that the proposed encoding algorithm can avoid the crosstalk artifacts and unrealistic fixed-spread assumption. Combining with the deconvolution-based objective function, the encoding algorithm can be implemented successfully without the true wavelet
CRediT authorship contribution statement
Qingchen Zhang: Methodology, Writing - original draft, Writing - review & editing. Weijian Mao: Supervision, Validation. Jinwei Fang: Software.
Acknowledgments
This work was jointly supported by State Key Research Development Program of China (2016YFC0601100, 2018YFC0310104), National Natural Science Foundation of China (U1562216) and Hubei Provincial Natural Science Foundation of China (2018CFB355).
References (48)
- et al.
A structural rank reduction operator for removing artifacts in least-squares reverse time migration
Comput. Geosci.
(2018) Two-dimensional frequency-domain visco-elastic full waveform inversion: Parallel algorithms, optimization and performance
Comput. Geosci.
(2011)- et al.
Time-domain seismic modeling in viscoelastic media for full waveform inversion on heterogeneous computing platforms with OpenCL
Comput. Geosci.
(2017) - et al.
Elastic full-waveform inversion based on GPU accelerated temporal fourth-order finite-difference approximation
Comput. Geosci.
(2020) - et al.
An application of vector wavefield decomposition to 3D elastic reverse time migration and field data test
Comput. Geosci.
(2019) - et al.
Acceleration of stable TTI P-wave reverse-time migration with GPUs
Comput. Geosci.
(2013) - et al.
A practical implementation of 3D TTI reverse time migration with multi-GPUs
Comput. Geosci.
(2017) - et al.
3D Hybrid-domain full waveform inversion on GPU
Comput. Geosci.
(2015) - et al.
3D Seismic reverse time migration on GPGPU
Comput. Geosci.
(2013) - et al.
3D Laplace-Domain full waveform inversion using a single GPU card
Comput. Geosci.
(2014)
An optimized parallelized SGFD modeling scheme for 3D seismic wave propagation
Comput. Geosci.
A new full waveform inversion method based on shifted correlation of the envelope and its implementation based on OPENCL
Comput. Geosci.
Incoherent dictionary learning for reducing crosstalk noise in least-squares reverse time migration
Comput. Geosci.
An efficient implementation of 3D high-resolution imaging for large-scale seismic data with GPU/CPU heterogeneous parallel computing
Comput. Geosci.
RTM Using effective boundary saving: A staggered grid GPU implementation
Comput. Geosci.
A recipe for practical full-waveform inversion in anisotropic media: An analytical parameter resolution study
Geophysics
Simultaneous-source full-wavefield inversion
Lead. Edge
An efficient multiscale method for time-domain waveform tomography
Geophysics
Geological structure guided well log interpolation for high-fidelity full waveform inversion
Geophys. J. Int.
A matrix-transform numerical solver for fractional Laplacian viscoacoustic wave equation
Geophysics
Source-independent time-domain waveform inversion using convolved wavefields: Application to the encoded multisource waveform inversion
Geophysics
Application of multi-source waveform inversion to marine streamer data using the global correlation norm
Geophys. Prospect.
Multi-source least-squares reverse time migration
Geophys. Prospect.
Simplified hybrid domain FWI method and GPU acceleration
Geophys. Prospect. Petrol.
Cited by (19)
Seeding alpine grasses in low altitude region increases global warming potential during early seedling growth
2024, Journal of Environmental ManagementMesoscale soil moisture survey by mobile cosmic-ray neutron sensor across various landscapes in the Heihe River Basin
2023, Research in Cold and Arid RegionsA biophysical model to simulate seasonal variations of soil respiration in agroecosystems in China
2023, Agricultural and Forest Meteorology