Selvauma/srf2-deepmd

SrF₂ — Li-Metal Anode Protective Coating (DeepMD)

What this is

DeepMD-kit potentials for SrF₂, studied as an F⁻-conducting protective coating on Li-metal anodes. Unlike the other families here, the mobile species of interest is the fluoride anion, not Li⁺ — the coating's job is to be an F⁻ conductor that physically/chemically shields the Li-metal surface.

Transport properties (from this project's own AIMD/MLMD)

	AIMD	MLMD (DeepMD)
D (F⁻, cm²/s)	4.8×10⁻¹¹	5.0×10⁻¹¹
Eₐ (eV)	0.57	0.58

Mechanical: E = 88 GPa, ν = 0.28, Ω = 19.5 ų/formula unit, ρ = 4.24 g/cm³.

Files & Validation

File	RMSE energy (eV/atom)	RMSE force (eV/Å)	Training steps	Size
model/srf2_aimd_run1.pb	0.00363	0.0865	1,000,000	207.1 MB
model/srf2_aimd_run2.pb	0.00608	0.0999	1,000,000	146.6 MB

RMSE values are validation-set (held-out) energy/force error, read directly from each run's DeepMD-kit lcurve.out at its final training step — not re-derived or estimated.

2 of 5 distinct potentials, both independent training runs against the same AIMD trajectory set (srf2_aimd_run1/run2) — selected because they're the only two with a preserved lcurve.out to validate against. The large-scale production potential actually deployed for the multi-nanosecond NVT/NPT interface runs (deepmd/n_opt/mlpot.pb) has no training log kept on disk, so it's excluded from this metrics-first selection — ask if you want it included anyway (unvalidated) or swapped in for one of these two. Several older potentials from earlier iterations of this system are archived/ superseded and also not included.

Training pipeline

AIMD (VASP, PBE) → DeepMD-kit train → freeze → compress. Produced by HPCA (github.com/selvachandrasekaranselvaraj/hpca).

Citation

Selva Chandrasekaran Selvaraj, University of Illinois Chicago.