al3ni-mace

<!-- Paste into the model card at https://huggingface.co/asiri1/al3ni-mace (README.md), after the model description and before any usage examples. Every number below is taken from results/VALIDATION_REPORT.md in the companion GitHub repository. Do not edit them here independently - if the report is regenerated, update this section from it, or it will drift. -->

These limits are not merely formal warnings. Beyond them, the model returns numbers that appear plausible but are incorrect, and nothing in the output alerts you to this.

Independent validation

An independent validation of al3ni_combined227_lora_v1 was run on separate hardware, OS and PyTorch version. The full report - 19 phases, from dataset integrity through MD stability to development applications - is here:

VALIDATION_REPORT.md

The datasets, splits, configs and status files it evaluates against are in the same repository: Asiri1108/ni-al-mlip.

The checkpoints in this repository are byte-identical to the ones the validation loaded; every SHA256 here was compared against the hash recorded during the run. To confirm before you rely on a result:

sha256sum al3ni_combined227_lora_v1.model
# e4fd54cc8a4a090fc9e32d6625269142824bfada8315433127b3aa3118c65cee

What it is validated for

• Relative energetics, forces and stresses of the five known Ni-Al phases within x_Ni in [0.25, 0.75], at strains up to about ±3%. Measured error there is ~0.4-1.3 meV/atom, force MAE ~0.001-0.003 eV/Å.
• Lattice parameters and equation of state (V0 to ≤ 0.2%).
• Formation energies and phase-stability ranking, using QE chemical potentials.
• Elastic constants for AlNi, AlNi3, Al3Ni, Al3Ni2 (cross-validated against LAMMPS).
• Phonon and dynamical-stability screening.

What it is NOT validated for, without further DFT

• Any composition outside x_Ni in [0.25, 0.75], including the pure elements. Pure Ni is wrong by +112.5 meV/atom.
• Strains beyond about ±4%, where error rises to 4.1 meV/atom mean, 9.7 maximum.
• Al3Ni at high expansion (+3.5 to +4.5%) at the 2.6183 meV/atom accuracy level.
• Quantitative Al3Ni5 shear or alpha-angle energetics.
• Radiation damage, collision cascades, or any close-approach regime.
• Defect, diffusion or creep studies relying on vacancy energies, until those are confirmed by DFT. No DFT ground truth exists for any vacancy quantity here.
• Novel structure prototypes, which the project's screening tool marks CAUTION by construction.

These bounds are not cautionary boilerplate. Outside them the model returns plausible-looking numbers that are wrong, and nothing in the output signals it.

Honest caveats

• No unopened seal exists — nothing in the validation is a blind test.
• Two quantities were found to be physically meaningless on first computation and were recomputed; both defects and their evidence are documented in the report's "Corrections applied during validation" section.
• MD trajectories are 5 ps. That is a stability and drift screen, not a converged transport or free-energy study.

Checkpoints in this repository

All 13 carry a recorded SHA256 and were loaded by the validation.

Base model: MACE-MATPES-PBE-0, verified byte-identical to the checkpoint recorded in the training provenance.

---

library_name: mace tags:

• mace
• interatomic-potential
• molecular-dynamics
• materials-science
• ni-al
• lora

---

Al3Ni / Ni-Al MACE potentials (LoRA fine-tunes of MACE-MATPES-PBE-0)

Three LoRA fine-tunes of the MACE-MATPES-PBE-0 foundation model on 227 QE/PBE configurations covering five Ni-Al intermetallic phases: Al3Ni, Al3Ni2, AlNi, Al3Ni5, AlNi3.

The three checkpoints are seed replicas: same training data, same hyperparameters, three different random seeds. They are published together because the reproducibility claim for this model is statistical, not bitwise (see below).

Files

SHA256

e4fd54cc8a4a090fc9e32d6625269142824bfada8315433127b3aa3118c65cee  al3ni_combined227_lora_v1.model
4328154a37f55c1ba99c209e978ad76a0f36e5f3d8a41b25c107b27efefe17af  al3ni_combined227_seed20260812_lora_v1.model
f676d87f8463a8ffcd4fa7d6b81796b5d42b89602267109010316cbdb4c97e6a  al3ni_combined227_seed20260813_lora_v1.model

The first hash is pinned inside the public configs/AL3NI_ROUND285_FINAL_UNSEALING_RESULT.txt, so the file you download can be confirmed to be the exact artifact named in the sealed record.

What the two seed replicas are for

They are not alternative production models — do not average them, and do not pick between them by which scores best on your system. They exist so that two specific claims can be checked by someone who does not have access to the original training environment:

1. The seed-noise floor. The three seeds agree to 0.078 meV/atom in overall relative-energy MAE, far inside the measured 0.43 meV/atom noise floor. Any effect smaller than that is seed noise, not signal — which is the standard a reader needs in order to judge which of the reported differences mean anything.

2. The 2.6183 meV/atom acceptance threshold. The project's locked acceptance bar is the maximum, across these three seeds, of each seed's own maximum |E_rel| error over the 19 reserved configurations. The per-seed maxima are 2.5232 / 2.4333 / 2.6183, and the same configuration (cfg060) anchors all three. Without the two replicas that number can only be asserted; with them it can be re-derived from published artifacts alone.

Reported relative-energy accuracy on the 227-config evaluation set:

E_rel = (E(config) − E(phase_relaxed))/N. Absolute MACE energies carry a model-dependent per-element offset, so only relative energies are comparable across models.

Usage

from mace.calculators import MACECalculator
from huggingface_hub import hf_hub_download

path = hf_hub_download("asiri1/al3ni-mace", "al3ni_combined227_lora_v1.model")
calc = MACECalculator(model_paths=path, device="cpu", default_dtype="float64")

atoms.calc = calc
atoms.get_potential_energy()

Use default_dtype="float64" — the reported errors are float64 numbers.

Scope — read before using

Fit for

• Relative energetics, forces and stresses of the five phases above, within x_Ni ∈ [0.25, 0.75], at strains up to about ±3%.
• Lattice parameters and equation of state (V₀ to ≤ 0.2%).
• Formation energies and phase-stability ranking, using QE chemical potentials.
• Elastic constants for AlNi, AlNi3, Al3Ni, Al3Ni2 (cross-validated against LAMMPS).
• Phonon and dynamical-stability screening.

Not fit for, without further DFT

• Any composition outside x_Ni ∈ [0.25, 0.75], including the pure elements.
• Strains beyond about ±4%, where error rises to 4.1 meV/atom mean, 9.7 maximum.
• Al3Ni at high expansion (+3.5 to +4.5%) at the 2.6183 meV/atom accuracy level.
• Quantitative Al3Ni5 shear or alpha-angle energetics.
• Radiation damage, collision cascades, or any close-approach regime.
• Defect, diffusion or creep studies relying on vacancy energies, until confirmed by DFT.
• Finite-temperature claims — the MD stability screen is not yet complete.

Reference lattice parameters

From this project's DFT-relaxed structures:

Reproducibility: statistical, not bitwise

Retraining from the public configs/al3ni_combined227_lora_v1.yaml is not expected to regenerate these files bitwise. The config sets a seed but no determinism controls, training ran on CUDA, and the toolchain (Python 3.12.3, torch 2.8.0+cu128, CUDA 12.8, RTX 4090) is pinned by documentation rather than by a container. This is normal for GPU training and is not a defect.

The claim that is supported is the one the three seeds above establish: independent training runs on identical data land within 0.078 meV/atom of each other.

Training data and code

Datasets, splits, configs and training scripts are public at Asiri1108/ni-al-mlip. Base model: ACEsuit/mace-foundations, release mace_matpes_0 (e618ad582b84239905b9c3b77ce6e9ce111b0ecd1533223a1a6aac7a696b8aa0).

Verifying on Windows: clone with git -c core.autocrlf=false clone …. With autocrlf enabled, git rewrites LF to CRLF on checkout and every text artifact's SHA256 will differ from the recorded value — 32 of 43 artifacts appear tampered with when they are not.