Allow non-conservative simulations without stress

[PicoCentauri]

This is a bit delicate. The problem seems to be that the vflag_global is kind of always set. I tried a minimal NVE simulation and a thermo_style where there is no pressure output. But even during this run, the flag is set to true.

The only workaround I see is to warn the user at init time and unfortunately silently ignore the pressure output during the run.

I am open to any better suggestions. @Luthaf or @HaoZeke.

[Luthaf]

If vflag_global is set, I think we should compute the virial. One solution if the model does not have non_conservative_stress would then be to fallback to backward propagation for this. So if the user sets non_conservative on, we try to use both nc_forces and nc_stress, falling back to conservative forces/stress if one of them is missing. If both are missing it should be an error IMO.

Alternatively, we could allow non_conservative forces/non_conservative stress in the lammps input to only use one of them.

I don't have strong feeling here, we can either automatically use autodiff when the non_conservative output does not exists; or force the user to tell us and have non_conservative on mean we use non-conservative for everything.

[PicoCentauri]

I think a fallback to autodiff and write a warning might not be read by the user. I am more in favour of raising an error if either non_conservative_forces or non_conservative_stress is not existing and forcing them to manually set the value of non_conservative. We could go for the options as you suggest @Luthaf:

1. non_conservative_forces - stress and forces are nc
2. non_conservative_force - forces are nc, stress is ad
3. non_conservative_stress - forces are ad, stress is nc

[Luthaf]

I'm fine with the error, but it seems that there is a typo (_ instead of ) in your example? I can see the following options in the LAMMPS input file:

• non_conservative off default, do not use any non-conservative output
• non_conservative on use both nc stress and forces
• non_conservative forces use nc forces and conservative stress
• non_conservative stress use nc stress and conservative forces

[PicoCentauri]

Yes sorry for the typo. Okay let's do this options.

[PicoCentauri]

I assume, we have to add the same logic to ASE once we merged this.

[Luthaf]

Yes, we can even start adding it now!

[Luthaf]

Looks good, just one question about the error message!

[Luthaf]

How does this looks like if the user made a typo in the variant name?

[PicoCentauri]

Arghh yes this might look. Let me try if we can make this better.

[PicoCentauri]

I changed the error message it now looks like:

ERROR on proc 0: output 'non_conservative_stress' not found in outputs
Exception raised from pick_output at /Users/runner/work/metatomic/metatomic/metatomic-torch/src/misc.cpp:188 (most recent call first):
frame #0: c10::Error::Error(c10::SourceLocation, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>) + 56 (0x105f3fd7c in libc10.dylib)
frame #1: metatomic_torch::pick_output(std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, c10::Dict<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, c10::intrusive_ptr<metatomic_torch::ModelOutputHolder, c10::detail::intrusive_target_default_null_type<metatomic_torch::ModelOutputHolder>>>, std::__1::optional<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>>) + 1548 (0x1064189e0 in libmetatomic_torch.dylib)
frame #2: LAMMPS_NS::PairMetatomic::settings(int, char**) + 4152 (0x104e32df8 in lmp)
frame #3: LAMMPS_NS::Input::pair_style() + 1460 (0x104b2ef78 in lmp)
frame #4: LAMMPS_NS::Input::execute_command() + 2380 (0x104b26b08 in lmp)
frame #5: LAMMPS_NS::Input::file() + 768 (0x104b259c0 in lmp)
frame #6: main + 80 (0x10490af98 in lmp)
frame #7: start + 6992 (0x18925fda4 in dyld)
Failed to select 'non_conservative_stress' output. If the model does not support non-conservative stress, use 'non_conservative forces' or 'non_conservative off'. If the model provides multiple variants, select one with 'variant/non_conservative_stress <name>'. (src/ML-METATOMIC/pair_metatomic.cpp:328)

[HaoZeke]

that's nicer, but maybe a few line breaks would help?

[PicoCentauri]

I can do one before our extra message in the last line.