lyc-sse-deepmd

LYC — Doped Li₃YCl₆ Halide Solid-State Electrolyte (DeepMD)

What this is

2 of 24 DeepMD-kit interatomic potentials trained for Li₃YCl₆ (LYC), a halide Li-ion solid-state electrolyte — selected from the full doping study below by lowest validation force RMSE (see Files & Validation). The full study covers a systematic doping series on two independent sites:

• Anion-site (Cl → F) substitution — 4 fluorination levels (F08/F16/ F24/F32, i.e. ~12/23/35/46 F substitutions per cell), tested against an independently-trained baseline set (LYC1_* vs LYC_*).
• Cation-site (Y → M) substitution — single dopants In, Yb, Zr, Er, Hf at 25/50/75/100% Y-site occupancy, and all six pairwise co-doped combinations (In+Yb, In+Zr, In+Er, In+Hf, Yb+Zr, Yb+Er, Yb+Hf, Zr+Er, Zr+Hf, Er+Hf) at matched 12/25/38/50% levels each.

Each .pb file is a separate, composition-specific model — trained on AIMD data for that one doping level, not a single transferable potential across the whole composition space. Don't extrapolate one composition's model to another; that's exactly what this series exists to systematically compare instead of assume. The two included here (LYC1_pure, LYC1_F08) are both from the baseline LYC1_* validation series — the best-converged pair in the whole study — not the main doped-composition series shown in the map below; ask if you'd rather have a doped composition (e.g. an In/Yb/Zr/Er/Hf variant) swapped in — the other 22 potentials still exist on disk.

Map shows the full 24-composition study for context — only the 2 above are included in this repo.

Why doped LYC

Li₃YCl₆ is a moisture-tolerant halide SSE candidate. Doping the Y or Cl sublattice is a standard lever for tuning Li⁺ vacancy concentration and migration-barrier landscape without changing the parent structure — this series exists to map how each dopant/level shifts ionic transport, screened at DFT cost via AIMD and then scaled to long-timescale MLMD with these potentials.

Training pipeline (per composition)

AIMD (VASP, PBE) at multiple temperatures → DeepMD-kit dp train on energies/forces/virial → dp freeze → dp compress (→ this .pb). Produced by the HPCA orchestration platform (github.com/selvachandrasekaranselvaraj/hpca).

Files & Validation

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.

Intended use / limitations

• Composition-specific: use the .pb matching your target doping level.
• Not validated for compositions or temperatures outside the AIMD training window for that composition.
• Research software / research potentials: validate before relying on results, same caveat as the orchestration platform that produced them.

Citation

Selva Chandrasekaran Selvaraj, University of Illinois Chicago.