software architecture around shenfun? about shenfun HOT 2 OPEN

I think your question is a little too broad to spark any new ideas. A solver library can be created in a lot of different ways. SpectralDNS is one library that initially only contained Navier-Stokes, and then grew from that. I'm sure it's not the ideal library, but it seems to be working:-) One thing is that I have considered rewriting it with a nicer interface to integrators, as I don't really like the way it is now. The integrators that are in Shenfun I find nicer. Did you look at those yet? I also want to add some adaptive integrators here eventually. The good thing with these is that they separate the right hand side of the equations into linear and nonlinear parts completely generically.

When it comes to gracefully switching between bases I'm not so sure this should be possible. If you have a solution in a Chebyshev basis and want it in Legendre, then you need to make a projection. There's really no other switch allowed. One thing that actually is allowed is to switch directly from a Dirichlet basis to a regular Chebyshev basis, because these bases can use the same quadrature points in physical space. There is no simple mechanism implemented to do this switch yet though. You need to create both bases and then copy the data from one physical array in one basis, to an array in the other one.

One option is to implement solvers in spectralDNS in the current framework. Have you thought about that possibility?

from shenfun.

yeah, I agree, this issue is a bit vague. It did steered some useful replies from your part though ;)
I'll look at integrators more closely, this looks interesting.

Regarding the bases switching issue, I am afraid I was not accurate enough in my initial post.
In a most simple situation, imagine we want to write a solver for an elliptic problem and we would like to let the basis choice as an option to the user. My intuition is that this may have an important impact on the architecture of the library but it's tough for me to estimate what this would be precisely.

I like your last suggestion about trying spectralDNS architecture.
Would it be difficult to build a minimalist version of spectralDNS?

from shenfun.