[Brean0]

Summary

PairPotentialModel adds each weighted pair energy to the system once and to each endpoint in full. As a result, the sum of energies is twice the returned system energy, for both full and half neighbor lists.

For two isolated soft-sphere atoms at distance 0.5, with epsilon=sigma=1 and alpha=2, the system energy is 0.125 but the atom energies are [0.125, 0.125]. Allocate half of the weighted pair energy to each endpoint, giving [0.0625, 0.0625] and preserving the system total of 0.125.

Reproduction

python -m pytest -q tests/models/test_pair_potential.py -k 'isolated_pair_atom_energies or batched_atom_energies_unequal_system_sizes'

The four isolated-pair cases and four unequal-batch cases fail at the expected per-atom energy assertion with baseline accumulation and pass with the fix. Both fixtures cover full and half lists with torch_nl_n2 and the default torchsim_nl backend on CPU float64. The unequal batch has three atoms in one system and two in the other; its expected atom energies are [0.0625, 0.125, 0.0625, 0.0625, 0.0625] and its system totals are [0.25, 0.125]. Each system's atom energies must sum to its own returned total.

Validation

• python -m pytest -q tests/models/test_pair_potential.py tests/models/test_soft_sphere.py tests/models/test_lennard_jones.py: 59 passed after adding the unequal-batch regression.
• Repository CI command python -m pytest --cov=torch_sim --cov-report=xml: exited before collection because the prepared environment does not have pytest-cov (--cov arguments were unrecognized).
• Final repository-wide python -m pytest: 1011 passed, 208 skipped, exit 0. Each final run includes all regression additions. Three PyTables UnclosedFileWarning messages followed the summary during trajectory-file cleanup.
• Ruff 0.17.1 lint and format checks on both changed files: passed.
• git diff --check: clean.

Scope and evidence

This change is limited to per-atom energy allocation and regressions. It leaves the system energy, force accumulation, stress accumulation, and neighbor-list construction unchanged. GPU and long-running simulation validation were not performed.

The F002 evidence and native ASE control documents the original invariant failure.

Checklist

• Docstrings have been added in the Google docstring format.
• Ruff lint and format checks pass.
• Regression tests have been added for the fix.

[lil-lon]

The fix looks good, but could you please simplify tests as suggested?

[lil-lon]

This path is covered by test_batched_atom_energies_unequal_system_sizes, whose second system is the same isolated pair, so I think you can delete this test.

[lil-lon]

The neighbor-list backend parametrization is unnecessary, since the neighbor-list backends are tested separately.
Keeping this test parametrized only over reduce_to_half_list should be enough.