Supervised neural network for surface waves data-driven Vs-model prediction

Both deterministic and probabilistic approaches to solve FWI usually lack of prior information that may serve as initial models or constraints for their application. Moreover, these methodologies are computationally expensive mainly because of the high number of forward modeling computations, especially for the elastic FWI case. We proposed to train a fully connected convolutional neural network to approximate the inverse operator. We compressed both surface-waves data and Vs-model into the Discrete Cosine Transform (DCT) domain, which served as input and output, respectively for the supervised training. we carefully selected the most common features of near-surface geology, such as landslide, sinkholes, stratification, layer displacements, and landfills. From these model bases we generated 10000 datasets assuming random realizations with different covariances. We demonstrated the neural network's ability to accurately predict intricate near-surface features characterized by strong lateral and vertical velocity variations. We hypothesize that the models generated by the neural network could be adequately reliable to serve as initial models for local inversion approaches, potentially facilitating the attainment of a global minimum. Additionally, these NN predicted models can offer valuable prior information in practical situations where we lack geological information.