Physical units in programming languages

Posted on 3 April 2022
Table of Contents:

Long time ago I read an article about Ada programming language, which included a demonstration showing how to use Ada’s strong type system for checking physical units of variables in a similar way how we rely on a compiler to check common data types. However the support of dimensional analysis and physical units in programming languages improved significantly in the meantime, and now one can find good support for it not just in “serious” languages like Ada. In this post we will see how to work with units on two simple examples, one for negative and the other for a positive use case, (re)implemented in Ada, F# and Python.


The example from the old article about Ada (2003, in Czech language) looked something like this:

type Meters is new Float;
type Meters_Squared is new Float;

-- Overloading multiplication operator on Meters data type so
-- that it will return Meters_Squared.
function "*" (Left, Right : Meters) return Meters_Squared is
  return Meters_Squared(Float(Left)*Float(Right));
  -- to avoid recursion, operands of the multiplication were
  -- overloaded to Float, this way we force compiler to use
  -- standard Float multiplication

  height : Meters := 10.0;
  width  : Meters := 15.0;
  surface_a : Meters_Squared;
  surface_b : Meters;
  surface_a := height*width; -- this is ok
  surface_b := height*width; -- causes compile time error

I recall that this really impressed me. But back then I wasn’t familiar with lot of programming languages, neither I was looking into this further. Unfortunately I also had a wrong impression for a while that such features are possible only with some strongly typed languages such as Ada. Which is not really the case as we will see later in this post.

When I revisited the example and wanted to try it out (GNU/Linux distributions like Fedora or Debian provides package for GNU Ada compiler GNAT), it turned out that the code needs few changes for it to actually work. Which in this particular case means that it will end up with a compile time error showing that the unit checking works as expected :-)

procedure Example1 is
  type Meters is new Float;
  type Meters_Squared is new Float;
  function "*" (Left, Right : Meters) return Meters_Squared is
    return Meters_Squared(Float(Left)*Float(Right));
  function "*" (Left, Right : Meters) return Meters is abstract;
  len_a : Meters := 10.0;
  len_b : Meters := 15.0;
  surface : Meters_Squared;
  len_sum : Meters;
  len_sum := len_a + len_b; -- ok
  surface := len_a * len_b; -- ok
  len_sum := len_a * len_b; -- invalid
end Example1;

Besides polishing required to turn it into a standalone Ada program, it was necessary to make function "*" (Left, Right : Meters) return Meters abstract to supresses this function to be inherited from multiplication on type Float. And then Ada compiler will really catch the error in dimension as expected, even though the error message doesn’t look very intuitive (tried with gcc-gnat-11.2.1-7 on Fedora 35):

$ gnatmake -q example1.adb
example1.adb:16:20: expected type "Meters" defined at line 2
example1.adb:16:20: found type "Meters" defined at line 2
gnatmake: "example1.adb" compilation error

Representing physical units in this way may not be straightforward nor practical. But for simple cases when we don’t need full dimensional analysis, this approach can work nicely, as we can see in strong typing example for handling meters and miles from course Introduction to Ada.

In a nice and extensive overview of dimensional analysis in programming languages, we will learn that this was already clear in the 80s, when dimensional analysis for Ada started to be investigated. Eg. N. H. Gehani in a paper from 1985 explains usage of a type system with operator overloading (as done in the example above) and concludes that it doesn’t generally work:

Derived types only partially solve the problem of detecting the inconsistent usage of objects; some valid usages of objects are also not allowed. Moreover, the solution is inelegant and inconvenient to use.

Besides further compiler research, focus was also given to design of libraries which defines data structures holding both value and it’s physical dimension along with functions operating on them. Something like this could be implemented in any language. That said guarantees given to a programmer and it’s cost will be constrained by chosen approach and design of a particular programming language. Sheer data structure library approach without using any additional language features (eg. type system extensions or compile-time macros) leads to impractical run-time checks. In an ideal case we want the dimension checks to performed during compilation to avoid additional run-time cost.

Nowadays Ada compiler GNAT provides native support for compile-time dimensional analysis. It uses aspect clauses from Ada 2012 standard to implement Dimension aspect which can be used to define dimensions for numeric types. GNAT library System.Dim.Mks uses it to define unit system according to SI standard. That said it puzzles me a bit that the library is called Mks, because MKS system defines just a subset of SI units.

When we rewrite the previous example utilizing this GNAT feature, we use unit symbol m for meters and dimension types Length and Area. All these types and symbols are defined in System.Dim.Mks library.

with System.Dim.Mks; use System.Dim.Mks;
procedure Example2 is
  len_a : Length := 10.0*m;
  len_b : Length := 15.0*m;
  surface : Area;
  len_sum : Length;
  len_sum := len_a + len_b; -- ok
  surface := len_a * len_b; -- ok
  len_sum := len_a * len_b; -- invalid
end Example2;

And as we can see, the compiler reports an error as expected, which is clearly explained (L is a dimension symbol for Length as defined in System.Dim.Mks library):

$ gnatmake -q -gnat2012 example2.adb
example2.adb:10:11: dimensions mismatch in assignment
example2.adb:10:11: left-hand side has dimension [L]
example2.adb:10:11: right-hand side has dimension [L**2]
gnatmake: "example2.adb" compilation error

Such support is a combination of direct implementation in a language (Dimension aspects) and a library (System.Dim.Mks) and having such support in a language and GNAT standard library makes it easier for dimensions to be part of a public API.

But let’s explore another more interesting (but still simple) example. Assume we have 1.2 volt 3000 mAh battery, and wonder how much energy in joules is stored there. Since we know that energy E equals charge Q times voltage U, this is no problem:

with System.Dim.Mks; use System.Dim.Mks;
with System.Dim.Mks_IO; use System.Dim.Mks_IO;
with Ada.Text_IO; use Ada.Text_IO;

procedure Example3 is
  U_b : Electric_Potential_Difference := 1.2*V;
  Q_b : Electric_Charge := 3000.0*mA*hour;
  E_b : Energy := U_b * Q_b;
  Put("charge = ");
  Put(Q_b, Aft => 2, Exp => 0);
  Put("energy = ");
  Put(E_b, Aft => 2, Exp => 0);
end Example3;

When we compile and run the program, we see that it provides expected results:

$ gnatmake -q -gnat2012 example3.adb
$ ./example3
charge = 10800.00 C
energy = 12960.00 J

Note that we see units for both values, since this information is available during runtime. That said we need to use Put from System.Dim.Mks_IO, standard functions from Ada.Text_IO are not able to work with dimensional quantities. And even though we haven’t explicitly done any unit conversion, we see that the charge is reported in coloumbs when printed to stdout, thanks to dimensional analysis and unit conversions performed during compilation. And last but not least we can be sure that we haven’t done any stupid mistake in units or computation itself, since the GNAT compiler haven’t reported any error.


Another language which is known for it’s good support for dimensional analysis is multi paradigm (functional/object oriented) language F#. I learned about it’s Units of Measure system in 2015 on Hacker News, but it was introduced back in 2008 and so it’s the first non experimental programming language with full builtin support for tracking of physical units.

Our previous simple example converted into F# would look like this:

[<Measure>] type m // declaration of unit of measure "m" representing meters

let len_a = 10.0<m>
let len_b = 15.0<m>
let len_sum : float<m>   = len_a + len_b // ok
let surface : float<m^2> = len_a * len_b // ok
let len_c   : float<m>   = len_a * len_b // invalid

In the first line we declare unit of measure m, which can be used in both numeric type declarations and numeric values. So float<m> is a type of float numbers of unit m, while float or float<1> denotes a unitless float value and 10.0<m> is a value of type float<m>. Note that we are unable to explicitly specify a dimension of the variable: we don’t state “this is a length specified as float meters”, but just “this is float meters”.

When we try to compile it, we will end up with expected unit of measure error:

$ dotnet run
/home/martin/projects/hello-fsharp/Program.fs(7,36): error FS0001: The unit of measure 'm' does not match the unit of measure 'm ^ 2' [/home/martin/projects/hello-fsharp/hello-fsharp.fsproj]

The build failed. Fix the build errors and run again.

This is very convenient way of representing units. That said, since we are working with standardized physical units, it’s crucial to not declare our own unit of measure for meters (as we did in the example above) and use standard definition from FSharp.Data.UnitSystems.SI library instead.

When we fix our simple example accordingly, we will end up with:

type [<Measure>] m = FSharp.Data.UnitSystems.SI.UnitNames.metre

let len_a = 10.0<m>
let len_b = 15.0<m>
let len_sum : float<m>   = len_a + len_b // ok
let surface : float<m^2> = len_a * len_b // ok
let len_c   : float<m>   = len_a * len_b // invalid

Now let’s have a look at the more interesting example. Note that because F# SI library doesn’t define any derived units for unit quantities (such as SI prefixed unit mA), we have to convert miliamps to amperes and hours to seconds themselves.

type [<Measure>] V = FSharp.Data.UnitSystems.SI.UnitNames.volt
type [<Measure>] A = FSharp.Data.UnitSystems.SI.UnitNames.ampere
type [<Measure>] s = FSharp.Data.UnitSystems.SI.UnitNames.second
type [<Measure>] J = FSharp.Data.UnitSystems.SI.UnitNames.joule

let u_b = 1.2<V>
let q_b = 3.0*3600.0<A*s>
let e_b : float<J> = u_b * q_b

printfn "charge = %.2f" q_b
printfn "energy = %.2f" e_b

And when we compile and run the program, we get the expected results:

$ dotnet run
charge = 10800.00
energy = 12960.00

Note that there is no way to get unit of a value since information about units is completelly lost during compilation. On one hand this is not necessary for dimensional analysis and validation itself, as it happens during compile time anyway. But on the other hand, it means you can’t implement any functionality which takes unit into account during runtime. To have the unit reported like in the previous Ada example, we would have to manually specify it like this:

printfn "energy = %.2f J" e_b

Which is obviously not verified by the compiler.


There are quite a few python libraries for physical unit handling. For the purpose of this blogpost, we are going to use Pint, which provides nice way to work with physical quantities with integration for packages like NumPy, Pandas and uncertainties. That said I don’t want to claim that Pint is the best option in all use cases, as I haven’t tested the alternatives.

We need to change our simple example a bit more for it to work in Python. Even though Python has a type system, it’s not possible to assign a type to a variable, so instead of our attempt to assign meters squared value into variable of length type, we will try to just sum meters with meters squared instead:

import pint

ureg = pint.UnitRegistry()

len_a = 10 * ureg.m
len_b = 15 * ureg.m
len_sum = len_a + len_b # ok
surface = len_a * len_b # ok
len_c = surface + len_b # invalid

So when we try to execute the script, we see an error as expected:

$ python
Traceback (most recent call last):
  File "/home/martin/projects/units/", line 9, in <module>
    len_c = surface + len_b # invalid
  File "/usr/lib/python3.10/site-packages/pint/", line 1079, in __add__
    return self._add_sub(other, operator.add)
  File "/usr/lib/python3.10/site-packages/pint/", line 115, in wrapped
    return f(self, *args, **kwargs)
  File "/usr/lib/python3.10/site-packages/pint/", line 989, in _add_sub
    raise DimensionalityError(
pint.errors.DimensionalityError: Cannot convert from 'meter ** 2' ([length] ** 2) to 'meter' ([length])

Of course another difference compared to previous examples in Ada or F# is that it’s a runtime error. For obvious reasons, pint is just a library providing classes like Quantity, which can be used to express and work with values and their physical units. That said if we really cared a lot about catching such errors during compile time, we won’t be using Python anyway.

To look at the more interesting example, we need to adapt it in a similar way as we did with the 1st python example. But here, we will explicitly convert the energy value e_b to joules (to make sure our assumption holds):

import pint

ureg = pint.UnitRegistry()

u_b = 1.2 * ureg.V
q_b = 3000 * ureg.mA * ureg.hour
e_b = (u_b * q_b).to(ureg.J)

print(f"charge {q_b}")
print(f"energy {e_b:.2f}")

So that we get this result:

$ ipython

In [1]: %run
charge 3000 hour * milliampere
energy 12960.00 joule

Note that charge q_b wasn’t converted into coulombs, since we haven’t performed any operation which would require that. But we can do such conversion explicitly as we did for energy e_b:

In [2]:
Out[2]: 10799.999999999998 <Unit('coulomb')>

Or we can check whether the quantity matches expected unit:

In [3]: q_b.check(ureg.C)
Out[3]: True


GNAT Ada or F# comes with very nice builtin support for work with physical quantities and dimensional analysis. And even though there are some limitations in both cases (which we haven’t looked into as as we just scratched the surface on few simple examples), overall both solutions are ready to be used for serious tasks. While run-time solutions can be implemented in any language, such solutions have additional cpu and memory overhead which makes them impractical. However for scripting languages or particular use cases, such cost can be justified, as we see for Python and Pint. If you work with physical quantities, usage of language feature and/or library for units it’s worth considering. And if you find this topic interesting, you may like already mentioned overview of Dimensional Analysis in Programming Languages, which covers this area in more detail and lists much more programming languages.


Links listed here are already present in text the post, but I’m highlighting them here again:

Do you have a comment or question? You can contact me.