Units through unitful?

[aramirezreyes]

Hi!

I am working on a couple packages. In particular this one. I was thinking on extracting my physicsfunctions.jl and adding them to a new package, but I discovered Thermodynamics.jl and now I am thinking on using it instead. One think that I am missing is support for units (in my package I am achieving that through Unitful.jl but this currently does not play nice with Thermodynamics.jl:

julia> using Thermodynamics, Unitful
julia> import CLIMAParameters as CP
julia> struct EarthParameterSet <: CP.AbstractEarthParameterSet end
julia> param_set = EarthParameterSet()
EarthParameterSet()
julia> saturation_vapor_pressure(param_set,300u"K",Liquid())
ERROR: MethodError: no method matching saturation_vapor_pressure(::EarthParameterSet, ::Quantity{Int64, 𝚯, Unitful.FreeUnits{(K,), 𝚯, nothing}}, ::Liquid)
Closest candidates are:

I usually use my functions as this:

julia> using TropicalCyclonePotentialIntensity, Unitful
julia> get_saturation_vapor_pressure(310.0u"K")
62.355321818976755 hPa

And I use units in my tests to ensure consistency and some other things. Considering that this package is being intended for wider use that only whithin CLIMA. Is accepting unitful units in the plans for Thermodynamics.jl? Would it be desirable?

[charleskawczynski]

Hi @aramirezreyes, good question. We put some effort into doing this a while back in ClimateMachine but I think that it didn't quite work for some reason.

We currently enforce pretty strict type diagonalization in Thermodynamics to avoid unwanted type promotions. However, if we fix #99, then we'll need to relax this restriction, and I do like the idea of having units. Perhaps we could do this, but it will require some work / thought.

[charleskawczynski]

I just realized that a nice alternative/compromise might be to add constructors that accept unitfull inputs, strip the units, compute necessary default constructor args, and re-wrap the default constructor args with the appropriate units.

one issue with this, however, is that we’ll need new types, since unitful stirs units in the type space. Alternatively we could use dynamic quantities (https://github.com/SymbolicML/DynamicQuantities.jl). However, that would require carrying around additional data, which. Perhaps we could do wrapping/unwrapping with that too?

wrapping and unwrapping kind of defeats the purpose of unitful internally, however, adding units generally speaking is not great for performance (or latency).