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:
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:
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.
Checkpoint
Role
al3ni_combined227_lora_v1
The production model. N=227, seed 20260811
al3ni_combined227_seed20260812/13_lora_v1
Cross-seed noise floor; derives the 2.6183 acceptance threshold
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
File
Seed
Role
al3ni_combined227_lora_v1.model
20260811
The production model. Use this one unless you specifically want a seed study. It is the artifact named in the project's sealed acceptance record and the one every reported validation number refers to.
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:
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.
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:
model
E_rel MAE (meV/atom)
E_rel RMSE
E_rel MAX
F MAE (eV/Å)
S MAE (GPa)
MACE-MATPES-PBE-0 (no fine-tuning)
4.1429
7.4763
53.5123
0.0188
0.8642
combined227 (seed 20260811)
0.9598
1.6103
9.7142
0.0012
0.2143
seed 20260812
1.0375
1.7653
10.3661
0.0013
0.2211
seed 20260813
1.0196
1.7211
10.1229
0.0013
0.2210
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.
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:
Phase
Structure
a (Å)
Notes
AlNi
B2 (CsCl-type)
~2.89
cubic, 2 atoms/cell
AlNi3
L1₂
~3.57
cubic, 4 atoms/cell
Al3Ni
orthorhombic
see MP mp-622209
16 atoms/cell
Al3Ni2
trigonal
see MP mp-1057
5 atoms/cell
Al3Ni5
orthorhombic
see MP mp-16514
8 atoms/cell
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.
