[Davis Rens (davrens101)]

Summary

Preserve nonperiodic coordinates when GraphConverter prepares mixed-PBC structures. Replace unconditional fractional-coordinate modulo with ASE's atoms.wrap(), matching the existing batch converter's treatment of periodic axes.

For a 10 Å cubic cell with pbc=[True, True, False], Si atoms at z = 1, 1.5 and 11.2 Å, and a 2 Å cutoff, the old converter moves the last atom to z = 1.2 Å. This creates six directed edges and six three-body entries instead of the physically expected two directed edges and no triplets. The same error occurs for a wire with pbc=[True, False, False].

The legacy converter remains the default calculator path (batch_converter=False, direct_graph=False). This therefore changes actual inference inputs, not just a deprecated helper in isolation.

Changes

• Use atoms.wrap() to respect the per-axis PBC mask.
• Add one parametrized regression for slab/wire geometries, checking preserved nonperiodic coordinates, periodic wrapping, exact edges and triplet count.
• Clarify the existing README limitations: geometric correctness does not establish non-bulk model accuracy.

Verification

Base: e59b088ee621cafe0df1b399a27299f3e9092ede.

• New regression against unchanged source: 2 failed, both because 11.2 Å became 1.2 Å.
• python -m pytest -q tests/datasets --disable-warnings: 11 passed after the fix.
• Real native three-body extension and both graph converters exercised on CPU: slab/wire topology now agrees with direct Cartesian separation and the batch converter. Fully periodic control retains six directed edges and six triplets; existing all-nonperiodic fallback tests pass.
• Actual MatterSimCalculator with a seeded, tiny initialized M3Gnet (no checkpoint download): the remote atom's spurious force disappears; the genuinely nearby geometry retains interactions. This is an execution smoke check, not a pretrained-model accuracy or performance claim.
• New regression file passes Black 23.3.0, isort 5.12.0 and Flake8 6.0.0. The converter already fails Black/isort and four Flake8 diagnostics at the unchanged base; this patch leaves those unrelated lines alone.

Local runtime: Windows, Python 3.12.10, PyTorch 2.15.0+cpu, torch-geometric 2.8.1, NumPy 2.5.4. Optional workflow dependencies were not all installed; no full-project-suite, CUDA, macOS or hosted-CI pass is claimed.

ASE may choose periodically equivalent representations near cell boundaries; bitwise identity of coordinates or image offsets is not promised. The reproduction uses a nonsingular cell and does not establish support for singular slab/wire cells.

AI assistance was used to investigate and draft this change; the regression and real graph/calculator execution were run locally. No independent human review is claimed.

[Davis Rens (davrens101)]

@microsoft-github-policy-service agree [company=]

[microsoft-github-policy-service]

Davis Rens (@davrens101) the command you issued was incorrect. Please try again.

Examples are:

@microsoft-github-policy-service agree

and

@microsoft-github-policy-service agree company="your company"