Figure 8.

Top panels show from left to right, the true velocity model taken from population V, the FNO inversion result using synthetic seismograms from one shot gather (one source and associated seismograms), and a comparison of the velocity profiles. The FNO network was trained using seismic data simulated in velocity models drawn from population T. The middle panels replicates the top panels except the FNO result is the sum of 20 different inversions for different sources in different locations, using the same model, i.e. stacking velocity models FNO predicted from 20 shot gathers with randomly located sources. The bottom panels show results after stacking 100 predicted velocity models. The right hand panels clearly show the convergence to an accurate velocity model after stacking multiple velocity model predictions derived from different shot gather FNO inputs, with the true model shown by the solid line and the prediction by the dashed line.