A new mathematical model for dispersion of Rayleigh wave and a machine learning based inversion solver

Jianxun Yang,Chen Xu,Ye Zhang

DOI: https://doi.org/10.48550/arXiv.2106.14025

2021-06-26

Abstract:In this work, by introducing the seismic impedance tensor we propose a new Rayleigh wave dispersion function in a homogeneous and layered medium of the Earth, which provides an efficient way to compute the dispersion curve -- a relation between the frequencies and the phase velocities. With this newly established forward model, based on the Mixture Density Networks (MDN) we develop a machine learning based inversion approach, named as FW-MDN, for the problem of estimating the S-wave velocity from the dispersion curves. The method FW-MDN deals with the non-uniqueness issue encountered in studies that invert dispersion curves for crust and upper mantle models and attains a satisfactory performance on the dataset with various noise structure. Numerical simulations are performed to show that the FW-MDN possesses the characteristics of easy calculation, efficient computation, and high precision for the model characterization.

Numerical Analysis

What problem does this paper attempt to address?

This paper attempts to solve two main problems:

1. **Establish a new Rayleigh wave dispersion function model**:
- By introducing the seismic impedance tensor, the author proposes a new function for Rayleigh wave dispersion in homogeneous and layered earth media. This new model provides an efficient method for calculating dispersion curves (the relationship between frequency and phase velocity).
- The formula is expressed as:
\[
\text{det}(\hat{Z}_0) = 0
\]
where \(\hat{Z}_0\) is the seismic impedance tensor at the surface.

2. **Solve the non - uniqueness problem of estimating S - wave velocity from dispersion curves**:
- When inverting Rayleigh wave dispersion curves to estimate crust and upper mantle models, the non - uniqueness problem is often encountered, that is, one dispersion curve may correspond to multiple different media models. This brings great difficulties to the inversion problem.
- To solve this problem, the author has developed a machine - learning method based on the Mixture Density Network (MDN), called FW - MDN, for estimating S - wave velocity. This method can handle the non - uniqueness problem and shows satisfactory performance on data sets containing various noise structures.

### Research background and significance

Rayleigh waves are surface waves. Their propagation speed is slower than that of P - waves and S - waves, so they are more sensitive to changes in the underground media. In addition, the energy of Rayleigh waves decays slowly and is easy to identify in seismic records. These characteristics make Rayleigh waves widely used in shallow geophysical exploration, deep seismology research and other scientific fields.

However, traditional dispersion curve inversion methods have problems of instability or high computational complexity. For this reason, the author proposes a new forward model and an MDN - based inversion solver, aiming to improve computational efficiency and accuracy while solving the non - uniqueness problem.

### Method overview

1. **New mathematical model**:
- By introducing the seismic impedance tensor, a new Rayleigh wave dispersion equation is established.
- Using this model, dispersion curves can be calculated efficiently.

2. **MDN - based inversion solver**:
- Use the Mixture Density Network (MDN) to handle non - uniqueness in the inversion problem.
- The FW - MDN model combines the prediction results of MDN and the forward model, and trains the network by optimizing the loss function.
- The loss function includes MDN loss, the distance between the predicted value and the true value, and regularization terms:
\[
\text{Loss} = -\sum_{j = 0}^{N - 1}\log(\hat{p}_{X|Y = y_j}(x_j))+\sum_{j = 0}^{N - 1}|\hat{y}_j - y_j|^2+\alpha_b\sum_{l = 0}^{B - 1}b_l^2+\alpha_w\sum_{l = 0}^{W - 1}w_l^2
\]

### Conclusion

This study proposes a new Rayleigh wave dispersion forward model and an MDN - based inversion solver, which solves the instability and non - uniqueness problems existing in traditional methods. The experimental results show that this method has significant advantages in computational efficiency and accuracy, and provides a new tool for estimating crust and upper mantle models.