In a former life I did HPC and I've cautioned often and strongly here against doing CFD because even for people who know what they're doing it's almost guaranteed to give pretty but misleading or meaningless results for people who don't know what they're doing.
That said, there is clearly a great deal of demand for CFD (not certain if that's real of just «it'd be cool..», but still).
One highly non-trivial thing that could be done is automatic / dynamic meshing with smart multi-physics to automagically figure out what regime a particular part of the simulation should be in and adjust accordingly.
Doing this would be a huge undertaking and I remain skeptical that it's worthwhile, but it'd be awesome.
If you wanted to go nuts, you could also add chemistry (multi-scale combustion physics) and if you wanted to go really insane, something like FLASH for nuclear CFD simulations (though you'd probably run into ITAR issues if you did that..)