Ludovic Räss

Thanks @ChrisRackauckas for reporting. Using ParallelStencil in inversion frameworks based on adjoint rules is definitively something that has potential ([here we derived the adjoint for multi-physics 3D problem](https://www.youtube.com/watch?v=1t1AKnnGRqA) - "not...

Thanks @ChrisRackauckas for your suggestions and further insights. I am still wondering which approach is generally followed under the hood by the workflows you describe to retrieve automatically the objects...

> Except that this doesn't work on the GPU What would be needed for it to run on GPU?

Hi @smartalecH, the issue you see passing a `struct` to a `@parallel` kernel relates to compliance limitation from CUDA.jl when it comes to passing `struct` to GPU kernels (from the...

> What do you guys typically do in these situations? Up to now, our approach was to do most of initialisation using native CPU arrays as you are referring to,...

Thanks @alexpattyn for your suggestion. I'd say the goal of the miniapps is exactly to feature domain specific examples on how to implement and leverage ParallelStencil (as this may often...

That sounds great @alexpattyn. Most valuable would certainly be to have time-varying source and PML within a single code. Output could be a double panel figure with forward simulation evolution...

Thanks @alexpattyn for sharing those details. Your absorbing boundaries implementation looks promising. As far as I remember from early experiments, the PML approach should ideally not generate any reflections (while...

Thanks for the update. I guess the making the absorbing layer gradually absorbing is indeed one way to go. I am still wondering wether this approach is then not rather...

That's interesting, thanks for reporting. I'd say that indeed, if the errors are sufficiently small we could go with it. Would be good, as you proposed, to quantify that briefly...