137 5. Full waveform inversion The FWI objective function decreases in the two steps (Figure 5.8). Here, we have shown all iterations, even though the L-BFGS algorithm does not consider all of them as real iterations. The reason is that each iteration has the same CPU cost. The objective function is normalized by its maximum value. After the first step (iteration 15), the objective function suddenly increases dramatically, even if the models are the same, simply because the frequency content of the data is modified. Most of the decrease is observed during the first iterations, while the final iterations are still important to retrieve the details in the model. Figure 5.8 Shape of the FWI objective function (normalised to 1) for the first 15 iterations (with fmin = 30 Hz and fmax = 60 Hz) and for the subsequent 30 iterations (with fmin = 30 Hz and fmax = 300 Hz). Figure 5.9 Observed data (left), initial data residuals (middle) and final data residuals (right), all represented at the same scale and for a shot at position x = 32 m. Here the pressure is displayed.
RkJQdWJsaXNoZXIy NjA3NzQ=