Engineered Atomic Structures — MLIP Zoo
Collection
DeepMD/MACE potentials and training data for SSE, Na-ion, Li-S, and cathode battery materials research. GPU-trained today! • 10 items • Updated • 1
MACE Fine-Tune: Cu / Sn / SrF₂ / Li
A MACE interatomic potential fine-tuned from the MACE-MP (Materials
Project) medium foundation model on Cu, Sn, SrF₂, and Li reference data —
built to support the LAMMPS-vs-NVIDIA-ALCHEMI benchmarking work
(alchemi-deepmd on GitHub: reproducible correctness/performance comparison
of LAMMPS+MACE, LAMMPS+DeepMD, ALCHEMI+MACE, ALCHEMI+DeepMD).
Status: in-progress, not converged. Training ran epoch 22→44 over ~12 hours on 2026-06-01, then stopped — no SLURM job is currently running, and MACE never produced a final SWA-averaged/compiled model (the marker it writes on completion). The two checkpoints here are real, genuine mid-training snapshots, not a finished release.
Training run
- Base checkpoint:
mace medium(MACE-MP / Materials Project foundation model,20231203mace128L1_epoch199model) — this is a fine-tune, not trained from scratch. - MACE version: 0.3.15
- Elements: Li, F, Cu, Sr, Sn (atomic numbers 3, 9, 29, 38, 50)
- Training set: 87,053 configurations (energy + forces)
- Validation set: 9,672 configurations
Files & Validation
| File | Epoch | RMSE energy (meV/atom) | RMSE force (meV/Å) |
|---|---|---|---|
model/mace_Cu_Sn_SrF2_Li_epoch44.pt |
44 | 88.98 | 81.83 |
model/mace_Cu_Sn_SrF2_Li_epoch6.pt |
6 | 89.09 | 89.43 |
RMSE values read directly from the run's own training log
(logs/mace_Cu_Sn_SrF2_Li_run-42.log) at each checkpoint's epoch — not
re-derived. Loss improved only modestly from epoch 22 (RMSE_F 84.30 meV/Å)
to epoch 44 (81.83 meV/Å) over ~12 hours — worth resuming and pushing
further before treating this as final, not concluding it's converged.
Related artifacts (not included here)
mace_small_lammps.pt / mace_small_mliap_nocueq.pt also exist in this
project's lammps_models/ directory — these are LAMMPS/MLIAP-deployment
exports of a generic "mace_small" model used as a baseline for the
speed-benchmarking work, not this Cu/Sn/SrF₂/Li fine-tune. Kept separate
here to avoid conflating a benchmarking baseline with an actual fine-tuned
result.
Training pipeline
MACE finetune mode against the MACE-MP medium foundation checkpoint,
GPU-trained (CUDA 12.1). Produced as part of the ALCHEMI/LAMMPS benchmarking
effort (github.com/selvachandrasekaranselvaraj/alchemi-deepmd).
Citation
Selva Chandrasekaran Selvaraj, University of Illinois Chicago.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support