objective function for gradient

[zhen0z]

Hi,

I'm trying to compute the gradient by defining a objective function like this:

function objective_function(model, state, Δt, step_i, forces)
    dif1 = JutulDarcy.compute_well_qoi(model, state, forces, :Producer1, SurfaceLiquidRateTarget) - obs_pro1[step_i]
    dif2 = JutulDarcy.compute_well_qoi(model, state, forces, :Producer2, SurfaceLiquidRateTarget) - obs_pro2[step_i]
    loss = sum(dif1.^2 + dif2.^2)
    return loss
end 

my observation data is:
obs_pro1= ws[:Producer1, :lrat]
obs_pro2= ws[:Producer2, :lrat]

However, the gradient computed based on this objective function seems not working quite well.

Therefore:
Is the function JutulDarcy.compute_well_qoi specifically designed for coarse and fine scale reservoir history matching?
Is it possible for me to access ws[:Producer1, :lrat] and ws[:Producer2, :lrat] through state in the objective function? ( I think JutulDarcy.compute_well_qoi is trying to complete this task, but it isn't giving me the right answer until now)

Thanks for your assistance!

Best regards,
Zhen

[moyner]

Hi,

The QOI code should be computing the point-wise data that ends up in ws[...] after some post-processing. I'm a bit surprised that the gradients are not accurate - could you share a small test case?

There is also a numerical gradient that uses the same interface - if the model is small, that is one way to verify if the gradients are accurate.

Best regards,
Olav

[zhen0z]

Hi Olav, Thanks for your clarification. The gradient computation seems to be working well. The issue was not with the gradient calculation but rather with my boundary conditions. I had set them as no-flow, which resulted in insufficient flow rate information to properly update the reservoir parameters. I apologize for the confusion caused earlier! Best regards, Zhen

…

Message ID: ***@***.***>