Cleanup

[cortner]

• Remove broken doc links in README and link to ACE1pack docs instead, tidy up the README
• probably remove examples -> again to ACE1pack? Some might go into ACE1x maybe...
• what to do about the scripts?

[cortner]

@tjjarvinen --- How much work would it be to create compatibility and regression tests so we can check for backward compatibility of all models, and run performance regression tests every time we tag a new version?

To me these seem to be things that fit into ACE1.jl and not ACE1pack.jl, but I'm happy to discuss.

[tjjarvinen]

It depends a lot on how the test are implemented. The best way is to implement them with meta programming, so that there is a loop over models that does all the tests. This way you can add more models by only adding model names to the loop.

[cortner]

I'm not sure how that would look like or how metaprogramming comes in. Why not just generate a representative list of models at different number of species, correlation orders, polynomial degrees etc and run the tests through all those. I was less worried about that part but more the infrastructrue / workflow.

[tjjarvinen]

The point is that you in general want to run the same tests for all the models, so you would need to repeat the same code many times. Also adding new models means a lot of work to implement the tests for that model.

This is an example what I have done with meta programming on this type of a case (now having different grid types)

gridtypes = [:ElementGridSymmetricBox, :CubicElementGrid]
for gt in gridtypes
     @testset "$(gt)" begin
            ceg = @eval $(gt)(5u"Å", 4, 16)
            r = position_operator(ceg)
            p = momentum_operator(ceg)
            x = r[1]
            @test unit(x) == u"bohr"
            @test x == r.operators[1]
            @test x != r.operators[2]
            @test x != r.operators[3]
            @test size(x) == size(r)
            @test_throws AssertionError r + p
            @test sin(1u"bohr^-1"*x).vals == sin.(x.vals)
            @test cos(1u"bohr^-1"*x).vals == cos.(x.vals)
            @test exp(1u"bohr^-1"*x).vals == exp.(x.vals)
            @test (x + x).vals == 2.0 .* x.vals
            @test (2x).vals ==  (x + x).vals
            @test (x^2).vals == x.vals.^2
            @test sqrt(x^2).vals ≈ abs.(x.vals)
            @test (x^2).vals == (x*x).vals
            @test unit(x^2) == u"bohr^2"
            @test unit(x+x) == u"bohr"
            @test unit(sqrt(x^2)) == u"bohr"
     end
end

[cortner]

ok, that's fine of course. I just think in our case, the meta-programming will be less important but the "list of models" will be constructed differently. Anyhow, those are just details I think.