vibe-qc/Documentation

Construction A qualification and operating limits

Construction A is qualified as an opt-in representation of ordinary periodic MACE for the operations below. It retains the mace-ccm-experimental name. It is not a modified finite-cluster potential, a newly trained model, or an isolated-defect method. The SECCM paper and its 0.18.x submission scope are unchanged. Source availability on main does not imply a release or backport.

Supported contract

Operation	Qualified scope
Energy and coordinate gradients	Neutral, multiplicity-one 3D periodic input; complete repeated input cells; all periodic image edges with unit weights
Strain, cell derivatives and stress	All nine strain components, fixed-fractional and fixed-Cartesian cell gradients, tensile-positive stress per input-cell volume
Fixed-cell relaxation	ASE BFGS, optional fixed atom indices, physical forces retained on all atoms
Joint atom/cell relaxation	ASE FrechetCellFilter with log strain scaled by input atom count; no fixed atoms
Recorded jobs	.out, .system, QVF; pinned construction/checkpoint; pending execution, replay and explicit failure status
Model/backend	Registered medium-mpa-0, mace-torch 0.3.16, CPU float64; tested with PyTorch 2.8.0 and ASE 3.29.0

Low-dimensional stress, charge/spin predictions, weighted neighborhoods, independently displaced replicas, Hessians, dynamics and embedding are outside this contract. Electronic properties are unavailable. Other backend versions must be qualified before admission because this adapter uses a private MACE batch/network interface. The original method="mace" route is separate.

Use one evaluator per sequential worker. Its model configuration is fixed. Each call rebuilds the discrete neighbor topology and differentiates coordinates and periodic shifts through the same scalar energy. At a cutoff crossing the adapter inherits the upstream model’s behavior; it introduces no boundary weights. The construction, normalization and derivatives are defined in the design note.

Calculation evidence

The #734-#736 evidence establishes primitive-cell energy/force/stress parity, all-component coordinate and cell finite differences, rotation/translation/ permutation checks, equivalent representatives and lattices, odd/even replicas, boundary ties, normalization and recorded-output round trips. Those passing calculations were reused. The #737/#738 extension adds bulk/defect size controls, optimization and resource measurements. Full inputs, output archives, exact source/native/checkpoint hashes and queue identities are retained on the tracker.

The extension used CPU float64 MPA-0 weights with SHA256 75428afe3a1d7d8062e19bcaabd5c433623cabf308242ec9fb493e38604fb638. The relaxation source landed in 6f8b46f937dbda30571a55b47dc8238573833c56. Execution used that Python implementation over a retained native core with SHA256 4bd9ac1a6465378e5cf895ff54d6a7db6704f9048c209e452c0fdd75d33505c9; it was not a fresh native build of all main changes. No C++ calculation algorithm was changed in this extension.

Frozen bounds were 1e-8 eV/input atom for energy parity, 1e-7 eV/angstrom for gradient parity, 1e-8 eV/angstrom cubed for stress parity, and 1e-5 eV/angstrom for central force differences at 1e-4 and 3e-5 angstrom. Across bulk/vacancy/EOS cases, maximum errors were respectively 4.84e-14, 4.99e-14, 1.40e-15 and 4.15e-8. The explicit expanded-cell oracle also passed. These are implementation checks against the same potential, not independent predictive-accuracy results.

The distorted two-atom diamond relaxation converged in four fixed-cell steps and nine joint atom/cell steps. Final residuals were 0.00966 and 0.02251 eV/angstrom against the declared 0.03 threshold. The latter is a combined cell-filter residual, not a raw stress. Fresh final energy, gradient and stress matched upstream MACE. Zero-step exhaustion and frozen atoms were exercised; unconverged jobs preserve the requested last geometry and do not claim success. Both converged optimization QVFs replayed through the CLI without another optimization step, retaining their pinned requests. Maximum replay errors were 3.03e-15 eV/atom, 1.65e-14 eV/angstrom and 2.65e-15 eV/angstrom cubed. A pending zero-step optimization settled its QVF as failed, preserving the submitted request and a valid schema.

Size and resource behavior

The following single-point jobs used skew strained diamond cells and cyclic replicas (2,1,1). Wall times include model loading and recorded output. Peak RSS is the high-water value of a fresh worker including its upstream comparison; the smallest vacancy worker also ran the expanded-cell and finite-difference controls. These are measurements, not scheduler memory estimates.

Input atoms	Lifted nodes	Directed edges	Cyclic job seconds	Worker peak RSS MiB
16 pristine	32	5056	7.02	1480
15 vacancy	30	4448	6.96	1650
54 pristine	108	17064	13.19	3019
53 vacancy	106	16432	12.70	2943
128 pristine	256	40448	23.84	6111
127 vacancy	254	39816	23.74	6095

The allocation guards of 4096 nodes and 1000000 directed edges do not guarantee a memory budget. The default (1,1,1) represents the same periodic potential with less replica work; it is not an isolated zero-image calculation. There is no speed or accuracy benefit inherent in increasing construction A replicas.

For the matched strained cells, the unrelaxed periodic vacancy excess E(N-1) - (N-1)*E(N)/N was 5.29022, 5.37202 and 5.39368 eV for 15, 53 and 127 atoms. The last increment is still 0.02165 eV. This size dependence is physical-input-cell dependence of the pretrained model, not cyclic-replica error. It is not evidence of isolated-defect convergence or defect accuracy. The model’s finite cutoff and multiple message-passing layers can mediate indirect defect-image interactions.

Independent plausibility screens and remaining science

Nine-point Birch-Murnaghan fits over lattice scales 0.96-1.04 gave:

Material	Model lattice / angstrom	Ambient reference / angstrom	Model bulk modulus / GPa	Reference / GPa
Diamond	3.57791	3.567	390.10	442.0
MgO	4.25605	4.21104	148.55	160.2

The preregistered screening windows were 3% for lattice and 20% for modulus. Both passed. Diamond reference values are the experimental comparison in Surratt et al. (1973), abstract. MgO uses Zha et al. (2000), the 300 K isothermal modulus and ambient density 3.585 g/cm cubed; the reference lattice is derived from that density and the rocksalt cell composition.

These broad screens compare a static learned potential with ambient measurements without thermal/zero-point corrections. In particular, the roughly 12% diamond modulus discrepancy is not chemical accuracy. There is no matched independent ab-initio force or vacancy-energy benchmark here. ORCA/CRYSTAL cannot establish identity to a pretrained neural network. Application-specific predictive claims need appropriate independent reference calculations and error thresholds; small same-model parity errors do not supply them.

The foundation-model study, Sections 4-5, describes training-domain, short-range and magnetic limitations. General materials coverage does not certify every defect or distorted environment. The modified construction B and dedicated-model track #739-#741 remain research, with no training campaign or new physical-model acceptance in this milestone.