Initial-model-free inversion of multimodal surface wave dispersion curves via implicit reparameterization

Surface wave dispersion inversion is essential for S-wave velocity estimation but suffers from severe initial-model dependence. Conventional mitigation strategies, such as global search or multi-start methods, incur prohibitive computational costs. We propose an initial-model-free inversion framework leveraging implicit reparameterization via a Multilayer Perceptron (MLP). By mapping depth to velocity within the network’s weight space, we transform the physical inversion into a parameter optimization problem. The inherent spectral bias of the MLP acts as an implicit regularizer, mitigating the need for accurate initial models. Furthermore, we integrate Monte Carlo Dropout (MCD) to achieve computationally efficient approximate uncertainty estimation in a single training process. Synthetic tests demonstrate stable convergence from homogeneous constant-velocity models—scenarios where gradient-based methods typically fail. Validation on the InterPACIFIC field dataset shows excellent agreement with borehole measurements. Finally, application to the Long Beach area confirms the method’s 3D scalability, successfully recovering complex geological structures from a constant-velocity initialization.