Learn to Invert: Surface Wave Inversion with Deep Neural Network

We propose a hybrid analytics and machine learning approach for large-scale surface wave inversion (SWI) for shear-wave velocities in the shallow overburden. A sparse grid of 1D velocity models are inverted using analytic optimization. Then, a deep neural network (DNN) with three hidden layers is trained using a spatially sparse subset of the data and non-linear inversion results. Finally, we use the DNN to predict the velocity model for the whole survey. This approach is demonstrated on a real high density land project. In comparison to the purely analytical approach, the hybrid analytic-ML method estimates a more reliable shear velocity model over the whole survey with significant reduction in computing time. We end with a discussion around the potential of this type of method for other geophysical inverse problems and seismic processing.