# Stork Land FWI benchmark

This model is expected to show that different Full Wave Form inversion methods will invert some parts of the model well, but will not do well with other parts of the model. The results simultaneously show the strength and weakness of FWI.

I welcome other algorithms to produce a better result and I expect others can. The 2D correct and starting model is attached. I propose this as a simple 2D standard. If your result is better, feel free to show my result as a comparison. Figure 1:Correct model

Figure 2: Starting model

A main weakness of FWI is that it has difficulty going very deep. The maximum offset is 10,000 meters and you see the inversion starts to get bad at a depth of 2500 meters. More iterations and using higher frequencies will improve the sharpness of the inversion and may extend the accuracy down a little further.

This inversion is 4-7.5 Hz. I ran only 23 iterations. Although this is a 2D model, since our code only runs in 3D, we ran this in 2.5D mode. Each iteration takes about 1 hour. For this synthetic, acoustic, isotropic 2D case, I expect one can do much better than this result with more iterations. But, for real data, elastic, anisotropic, 3D, I'm not sure we can do much better.

A good measure I like to track is velocity-RMS-change per data-RMS-reduction.

Note that I show two dramatically different results: one WITH data pre-conditioning and one WITHOUT data pre-conditioning. The data pre-conditioning dramatically helps the inversion by increasing the amplitude of deeper traveling energy. It does this by performing a FK-filter to reduce (~26db) but not eliminate the amplitude of slow horizontally travelling surface energy. The FK-filter is applied on the observed and synthetic data. This may be useful for land FWI.

For a tar file containing all these images and binary grids of the true model and the starting model click here.

Note that this test model is a slight modification of the earlier 2D model I distributed. The main change is that this model has a simpler near-surface.

Let me know if you have any questions.

Best regards,

Christof