Using convolutional neural networks to develop starting models for near-surface 2-D full waveform inversion

Non-invasive subsurface imaging using full waveform inversion (FWI) has the potential to fundamentally change engineering site characterization by enabling recovery of high resolution 2D/3D maps stiffness. Yet, accuracy FWI remains quite sensitive choice initial starting model due complexity and non-uniqueness inverse problem. In response, we present novel application convolutional neural networks (CNNs) transform an experimental seismic wavefield acquired a linear array surface sensors directly into robust for 2D FWI. We begin describing three key steps used developing CNN, which include: selection network architecture, development suitable training set, performance training. The ability trained CNN predict was compared against other commonly models classic near-surface problem; identification undulating, two-layer, soil-bedrock interface. developed during this study able complex images testing set from their wavefields with average mean absolute percent error 6%. When common approaches, approach produce smaller image misfits, both before after generalize were dissimilar ones upon it assessed more complex, three-layered model. While predictive slightly reduced, still achieve misfits comparable models. This demonstrates that CNNs have great as tool good ...