search

slimgroup / WISE.jl

WISE: full-Waveform variational Inference via Subsurface Extensions
https://doi.org/10.1190/geo2023-0744.1
MIT License
2 stars 0 forks source link

Replicating COMPASS results #6

Closed advaitb closed 4 months ago

advaitb commented 4 months ago

Hi @ziyiyin97,

I'm running into some issues with replicating the COMPASS results with a 1d gradient background. I'm getting very fuzzy posterior samples around the salt/seabed region as shown below. Any ideas on how to fix this or what's going wrong? I've trained it on 400 slices.

image

ziyiyin97 commented 4 months ago

When you say 1d gradient background, do you mean a velocity that monotonously increases along time? If that's the case, then your result would be worse. In the WISE paper, we used a 1d average of the compass dataset, which is a better background model so it produces better posterior samples.

On a side note: sorry for the delay, I am swamped by a few other things at this moment. I will try to push out the WISER code in a few days. Thanks in advance for your understanding.

advaitb commented 4 months ago

Interesting, maybe I'm getting this wrong but I used the 1d gradient and it looks closer to what you used in the paper in 2(b) ? If I use the 1d average I get a smoothed original image? Does this look correct to you?

Gradient: image Average: image

ziyiyin97 commented 4 months ago

1D gradient is indeed challenging because the travel time might be extremely off. I used all 1040 samples to compute the average and then smoothed it as indicated in my code. I suspect you use fewer samples? I don't expect structures in the averaged velocity model.

advaitb commented 4 months ago

Thanks, that worked and I was able to approximately replicate results from the paper!