Abstract:Full waveform inversion (FWI) infers the subsurface structure information from seismic waveform data by solving a non-convex optimization problem. Data-driven FWI has been increasingly studied with various neural network architectures to improve accuracy and computational efficiency. Nevertheless, the applicability of pre-trained neural networks is severely restricted by potential discrepancies between the source function used in the field survey and the one utilized during training. Here, we develop a Fourier-enhanced deep operator network (Fourier-DeepONet) for FWI with the generalization of seismic sources, including the frequencies and locations of sources. Specifically, we employ the Fourier neural operator as the decoder of DeepONet, and we utilize source parameters as one input of Fourier-DeepONet, facilitating the resolution of FWI with variable sources. To test Fourier-DeepONet, we develop three new and realistic FWI benchmark datasets (FWI-F, FWI-L, and FWI-FL) with varying source frequencies, locations, or both. Our experiments demonstrate that compared with existing data-driven FWI methods, Fourier-DeepONet obtains more accurate predictions of subsurface structures in a wide range of source parameters. Moreover, the proposed Fourier-DeepONet exhibits superior robustness when handling data with Gaussian noise or missing traces and sources with Gaussian noise, paving the way for more reliable and accurate subsurface imaging across diverse real conditions.
What problem does this paper attempt to address?
The problems that this paper attempts to solve mainly focus on two key challenges in full waveform inversion (FWI): generalization ability and robustness. Specifically:
1. **Generalization ability**: Existing data - driven FWI methods are severely limited in performance when dealing with different source functions (such as sources with varying frequencies and positions). This is because the source functions used in the training process may differ from those used in actual field investigations, resulting in the model being unable to adapt well to new data conditions. To overcome this challenge, the paper proposes a new model - Fourier - DeepONet. By introducing the Fourier neural operator as a decoder and using source parameters as one of the inputs, the generalization ability of the model for different source functions is improved.
2. **Robustness**: Data - driven FWI models are usually very sensitive to noise or outliers in the input data. This limits the reliability and accuracy of the model in practical applications. The paper verifies through experiments the performance of Fourier - DeepONet in handling data containing Gaussian noise and in cases of missing traces and sources, demonstrating that this model has stronger robustness.
### Main contributions of the paper
- **Model innovation**: Proposed Fourier - DeepONet, which combines the advantages of the Fourier neural operator and DeepONet, improving the generalization ability and robustness of the model.
- **Dataset development**: Developed three new benchmark datasets (FWI - F, FWI - L, and FWI - FL) that contain sources of different frequencies and positions for testing the generalization ability of the model.
- **Experimental verification**: By comparing with existing data - driven FWI methods (such as InversionNet and VelocityGAN), it shows that Fourier - DeepONet has more accurate prediction results and stronger robustness under multiple source parameter conditions.
### Paper structure
1. **Introduction**: Introduced the basic concepts of FWI and its importance in geophysics, as well as the challenges faced by current data - driven methods.
2. **Problem setting**: Described in detail the mathematical model of FWI and the forward modeling process, and introduced the datasets used.
3. **Method**: Introduced the architecture and working principle of the traditional DeepONet and its improved version, Fourier - DeepONet.
4. **Results**: Verified through experiments the performance of Fourier - DeepONet under different source frequencies, positions, and noise conditions, and compared it with the baseline model.
5. **Discussion and future directions**: Summarized the main findings of the research and discussed future research directions.
### Key formulas
- **Acoustic wave equation**:
\[
\nabla^2 p(x, z, t)-\frac{1}{c(x, z)^2} \frac{\partial^2 p(x, z, t)}{\partial t^2}=s(x, z, t)
\]
where \(p(x, z, t)\) is the pressure wave field, \(c(x, z)\) is the velocity map, and \(s(x, z, t)\) is the source function.
- **Ricker wavelet**:
\[
s_0(t; f)=(1 - 2\pi^2 f^2 t^2) e^{-\pi^2 f^2 t^2}
\]
where \(f\) is the frequency and \(t\) is the time.
- **Output of Fourier - DeepONet**:
\[
u(y)\approx G(v)(y)=\sum_{i = 1}^{r} b_i(v) t_i(y)+b_0
\]
where \(b_i(v)\) and \(t_i(y)\) are the outputs of the branch network and the trunk network respectively, and \(b_0\) is the bias term.
Through these innovations and experiments, the paper shows the significant advantages of Fourier - DeepONet in improving the generalization ability and robustness of FWI.