Using custom dimensions from DynamicQuantities

[HenryDane]

Hello,

I'm trying to do equation discovery in a dimensionally consistent way but I'm having a hard time of it. Many of my quantities are either number concentration (#/kg) or mass mixing ratio (kg of A/kg of B). I can not figure out how to express these units using the PySR utility.

I've tried switching to the SymbolicRegression.jl library and then implementing my own set of dimensions (following the documentation for DynamicQuantities) as show here:

# ...

# create our dimensions
struct ExtendedDimensions{R} <: AbstractDimensions{R}
    length::R
    mass::R
    time::R
    current::R
    temperature::R
    luminosity::R
    amount::R
    angle::R
    mass_h2o::R
    mass_air::R
    count::R
end

# ... 

const mmr   = Quantity(1.0, ExtendedDimensions(mass_h2o=1, mass_air=-1))
const nconc = Quantity(1.0, ExtendedDimensions(count=1, mass_air=-1))
const dens  = Quantity(1.0, ExtendedDimensions(mass_air=1, length=-3))

X = (; A=data.A * mmr, B=data.B * nconc)
y = data.C * mmr

# create model
model = SRRegressor(
	binary_operators=[*,^,/],
	dimensional_constraint_penalty=10^5,
	nested_constraints=[(^) => [(^) => 0]],
	constraints=[(^) => (-1,1)]
)

# create machine
mach = machine(model, X, y)

# do the fitting
fit!(mach)

r = report(mach)

but I get this error when I run the script:

ERROR: LoadError: Cannot create a dimensions of `Dimensions{DynamicQuantities.FixedRational{Int32, 25200}}` from `ExtendedDimensions{DynamicQuantities.FixedRational{Int32, 25200}}`. Please write a custom method for construction.
Stacktrace:
  [1] Dimensions{DynamicQuantities.FixedRational{Int32, 25200}}(d::ExtendedDimensions{DynamicQuantities.FixedRational{Int32, 25200}})
    @ DynamicQuantities ~/.julia/packages/DynamicQuantities/5QflN/src/types.jl:118
  [2] convert(::Type{Dimensions{DynamicQuantities.FixedRational{Int32, 25200}}}, d::ExtendedDimensions{DynamicQuantities.FixedRational{Int32, 25200}})
    @ DynamicQuantities ~/.julia/packages/DynamicQuantities/5QflN/src/utils.jl:323
  [3] convert(::Type{Quantity{Float64, Dimensions{DynamicQuantities.FixedRational{Int32, 25200}}}}, q::Quantity{Float64, ExtendedDimensions{DynamicQuantities.FixedRational{Int32, 25200}}})
    @ DynamicQuantities ~/.julia/packages/DynamicQuantities/5QflN/src/utils.jl:314
  [4] get_units(::Type{Float64}, ::Type{Dimensions{DynamicQuantities.FixedRational{Int32, 25200}}}, x::ExtendedDimensions{DynamicQuantities.FixedRational{Int32, 25200}}, ::Function)
    @ SymbolicRegression.InterfaceDynamicQuantitiesModule ~/.julia/packages/SymbolicRegression/RwU5c/src/InterfaceDynamicQuantities.jl:39
  [5] get_si_units(::Type{Float64}, units::ExtendedDimensions{DynamicQuantities.FixedRational{Int32, 25200}})

I'm very new to Julia, but it seems to me like the SRRegressor needs to "cast" a derived type (ExtendedDimensions) to a "base" type (Dimensions), and as I understand that, it is not something that can be done.

Does anyone have any advice? Is it possible to use a custom version of Dimensions?

(I'm asking here since it did not seem reasonable to open an issue on SymbolicRegression.jl's github for this, and the forums link there sent me here).

[MilesCranmer]

Hi @HenryDane,

Thanks for the post! Do you want to post this is as an issue on SymbolicRegression.jl? Custom AbstractDimensions have actually not yet been set up within SymbolicRegression – simply because I didn't know anybody would want to customize this :). But that sounds like a cool use-case, and it should hopefully be fairly easy to add this.

Cheers,
Miles

[HenryDane]

I've opened an issue here: MilesCranmer/SymbolicRegression.jl#335

[HenryDane]

Sure -- I'll open an issue on the SymbolicRegression.jl github.

[MilesCranmer]

P.S., as a quick hack, you could treat the other existing units as the units you are interested in. So, e.g., luminosity (cd) could be mass_h20.

[HenryDane]

That is a great idea, I'm not sure how I didn't think of that. I'll give it a try!