is it possible to call JUDI in python?

[zhen0z]

Hi,

I'm trying to call JUDI from python, is it possible to use pycall to call JUDI from julia? Thanks.

[mloubout]

I mean, in theory, it should be doable using pyjulia, but that would be quite complicated since JUDI already calls Python for the devito backend. You'd end up having "python call julia call python" which doesn't seem like a good idea.

[zhen0z]

Thanks for the explanation. It does seem difficult as I always run into memory conflict issues. Mathias Louboutin ***@***.***> 于2025年4月16日周三 13:10写道：

…

I mean, in theory, it should be doable using pyjulia <https://github.com/JuliaPy/pyjulia>, but that would be quite complicated since JUDI already calls Python for the devito backend. You'd end up having "python call julia call python" which doesn't seem like a good idea. — Reply to this email directly, view it on GitHub <#286 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BIOR2UQYXER63MCPI2IRGGD2ZZCETAVCNFSM6AAAAAB3IBSGEWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDQMBZGM4DKNBZHE> . You are receiving this because you authored the thread.Message ID: ***@***.***> *mloubout* left a comment (slimgroup/JUDI.jl#286) <#286 (comment)> I mean, in theory, it should be doable using pyjulia <https://github.com/JuliaPy/pyjulia>, but that would be quite complicated since JUDI already calls Python for the devito backend. You'd end up having "python call julia call python" which doesn't seem like a good idea. — Reply to this email directly, view it on GitHub <#286 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BIOR2UQYXER63MCPI2IRGGD2ZZCETAVCNFSM6AAAAAB3IBSGEWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDQMBZGM4DKNBZHE> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[mloubout]

Why do you want to call judi from python?

[zhen0z]

Basically, I try to use Judi as a forward function to compute the output and the corresponding gradient, and then use inversion code written in python to perform inversion. Do you have any suggestions about this?

[mloubout]

If your main code is in python I would probably do everything in python since devito (the actual propagators) are in python. JUDI's python backend should be usable directly from python you'll just need to read through it to figure out the input/outputs it expects

[zhen0z]

Thanks for the suggestions. Is it possible to link the fluid flow with wave equation and compute the gradient of the waveform with respect to the permeability in devito? I noticed Judi with JutulDarcyRules can do it, but it seems difficult to link with python.

[mloubout]

JUDi and Jutul are two separate software programs that handle different physics; they are linked together in our framework so that the chain rule works properly but the jutul part is not using devito. So if you have your own fluid solver with the sensitivities then yes you can completely do that you just need to add the proper rules and workflow so that the gradients propagate through the different physics.

[zhen0z]

Thanks for the explanation, it's very helpful. Let's say I can use Jutul in julia to compute the gradient of
d (real saturation - predicted saturation)^2 / d permeability

Also in python, I can get the gradient of
d (real waveform- predicted waveform)^2 / d velocity

plus the gradient of
d velocity / d saturaiton

Can I link them together to compute the gradient of
d (real waveform- predicted waveform)^2 / d permeability ?

I'm a bit confuse about how to achieve this?