Comparison of data-driven and physics-driven surface wave inversion

In near-surface investigations, the advent of massive seismic data has ushered in the application of deep learning (DL) techniques for surface wave inversion to attain the shear-wave velocity (Vs). While the efficiency of DL inversion surpasses that of classic physics-driven methods, its broader attributes remain underexplored. Our study delves into a comparative analysis of DL inversion versus physics-driven inversion, focusing on three key aspects: anti-noise ability, stability, and generalization.

In numerical experiments, we employ the neighborhood algorithm (NA) (Wathelet, 2008) as a representative of physics-driven inversion, and a convolutional neural network (CNN) constructed for near-surface investigations (Chen et al., 2022) as a representative of data-driven inversion. In addition to comparing the two methods, we also explore the characteristics of joint inversion using Rayleigh-wave dispersion curves (DCs) and Love-wave DCs in the three above aspects. To quantitatively evaluate inversion results, we calculate the root mean square error and relative error of both DCs and Vs. The assessment of anti-noise performance involves applying NA and CNN to DCs with varying noise levels. To gauge stability, we introduce errors in compressional-wave velocity (Vp) and density, examining their effects on inversion precision. Lastly, to assess generalization, we use NA and CNN to invert DCs whose Vs exceeds the range of the training dataset by different percentages.

Our findings reveal that DL inversion has a higher anti-noise ability compared with NA. Both methods demonstrate high stability, with errors in Vp and density exerting a slight impact on inversion results, aligning with surface wave inversion characteristics. Compared with physics-driven inversion, generalization is a unique feature of data-driven inversion. The experimental results indicate that the CNN can predict Vs models that are not included in the training dataset although this ability is somewhat limited. Furthermore, like physics-driven inversion, joint inversion enhances all three examined aspects for data-driven inversion. This analysis of characteristics can guide the selection of inversion methods for surface wave applications in near-surface investigations.

 

References:

• Wathelet M., "An improved neighborhood algorithm: parameter conditions and dynamic scaling," Geophysical Research Letters, vol. 35, no. 9, pp. 2008, doi: 10.1029/2008GL033256.
• Chen X., Xia J., Pang J., Zhou C., and Mi B., "Deep learning inversion of Rayleigh-wave dispersion curves with geological constraints for near-surface investigations," Geophysical Journal International, vol. 231, no. 1, pp. 1-14, 2022, doi: 10.1093/gji/ggac171.