Why Choose Deep Learning for Surface-Wave Inversion

Abstract

With the rapid advancement of artificial intelligence, deep-learning-based inversion frameworks are increasingly being adopted to tackle the challenges associated with surface-wave dispersion curve (DC) inversion. Compared with classical model-driven methods, the deep-learning-based inversion is known for its higher efficiency and independence from the initial model. Existing researches, however, have focused on algorithm design and case applications. The reforms that deep learning techniques can bring to inversion need further exploration. Therefore, we explored the anti-noise ability, stability, performance in joint inversion scenarios, and generalization ability of deep-learning-based inversions. For the first three characteristics, we select a published neural network and the neighborhood algorithm as representatives of deep-learning-based and model-driven inversions, respectively, to compare the corresponding performance of these two methods. The comparative tests and statistical analyses reveal that deep-learning-based inversion exhibits superior anti-noise ability and stability, but shows limited improvement in joint inversion performance. And the statistical results from tests for generalization ability show that the trained neural network can predict the shear-wave velocity (Vs) model whose Vs oversteps the model space of training dataset within 20%. In particular, we discover that the generalization ability is positively correlated with the prediction precision of Vs. This analysis provides valuable insights for choosing appropriate inversion methods and contributes to a deeper understanding of deep-learning-based inversions.