KyuJung Jun, Grace Wei, Xiaochen Yang, Yu Chen, and Gerbrand Ceder published a beautiful study last year in Matter on the LiMXCl4 superionic conductor family. The paper introduces the "soft cradle effect": the idea that framework flexibility of 1D [MO₂Cl₄]³⁻ octahedral chains, where O²⁻ bridges metal centers along the c-axis, creates an optimized energy landscape for Li⁺ migration. LiTaOCl₄ achieves 12.4 mS/cm at room temperature with an activation energy of just 0.23 eV.
I wanted to see how well Ouro's ML infrastructure handles these structures. Jun is a co-author on CHGNet (the model we use for charge-informed atomistic modeling), so this felt like a natural test case.
I constructed 6 CIFs from the LiMXCl4 family, varying the metal (Nb/Ta), the bridge anion (O/S), and the halide (Cl/Br):
LiNbOCl₄ — the primary experimental compound (10.4 mS/cm)
LiTaOCl₄ — highest conductivity (12.4 mS/cm)
LiNbSCl₄ — S-bridged variant
LiTaSCl₄ — S + Ta variant
LiNbOBr₄ — Br-substituted
LiTaOBr₄ — Br + Ta
All structures were built in I4/mmm (the high-symmetry parent of the experimentally reported I4/m), with bond lengths calibrated to typical M⁵⁺-O²⁻ (~1.90 Å), M⁵⁺-Cl⁻ (~2.35 Å), M⁵⁺-S²⁻ (~2.40 Å), and M⁵⁺-Br⁻ (~2.50 Å) distances. Each was relaxed through Orb v3 (conservative, inf cutoff, MPA) with full cell+ionic optimization, then checked against the Materials Project convex hull.
Compound | Input SG | Output SG | ΔE (eV) | ΔE/atom | Steps | Verdict |
|---|---|---|---|---|---|---|
LiNbOCl₄ | I4/mmm | I4/mmm | -0.317 | -0.020 | 12 | ✅ Preserved |
LiTaOCl₄ | I4/mmm | P1 | -1.594 | -0.100 | 93 | ❌ Collapsed |
LiNbSCl₄ | I4/mmm | P1 | -9.909 | -0.619 | 67 | ❌ Collapsed |
LiTaSCl₄ | I4/mmm | P1 | -9.681 | -0.605 | 49 | ❌ Collapsed |
LiNbOBr₄ | I4/mmm | I4/mmm | -0.585 | -0.037 | 12 | ✅ Preserved |
LiTaOBr₄ | I4/mmm | I4/mmm | -0.364 | -0.023 | 10 | ✅ Preserved |
Three structures survived, three collapsed to triclinic P1. The pattern is striking.
LiNbOCl₄ survives. LiTaOCl₄ does not. Despite nearly identical chemistry (Ta sits directly below Nb in Group 5), the Ta variant collapses with a 1.594 eV energy drop over 93 optimization steps, while the Nb variant relaxes cleanly in 12 steps with only 0.317 eV of energy relief. This matters because LiTaOCl₄ is the experimentally best conductor in the family.
The S-bridged compounds collapse catastrophically. Both LiNbSCl₄ and LiTaSCl₄ show ~9.7-9.9 eV energy drops, indicating the S-bridged chain structure I constructed is very far from what Orb v3 considers a stable configuration. The S²⁻ anion is significantly larger than O²⁻ (ionic radius 1.84 vs 1.40 Å), and the resulting chain geometry likely requires fundamentally different lattice parameters than my estimates. The "soft cradle" may be too soft for Orb v3 to handle.
The Br variants are the most robust. Both LiNbOBr₄ and LiTaOBr₄ preserve symmetry with small energy changes and fast convergence (10-12 steps). The larger Br⁻ anions may fill the unit cell more completely, giving Orb v3 a clearer energy landscape.
For the three structures that preserved symmetry, I ran the Materials Project convex hull check:
Compound | E_above_hull (eV/atom) | Stable? | Formation E (eV/atom) |
|---|---|---|---|
LiNbOCl₄ | 0.447 | No | -1.105 |
LiNbOBr₄ | 0.466 | No | -0.962 |
LiTaOBr₄ | 0.491 | No | -1.048 |
All three sit 0.45-0.49 eV/atom above the convex hull. This is consistent with the known metastability of oxyhalide solid electrolytes. These materials are synthesized by mechanochemical ball milling, not conventional solid-state routes, precisely because they are kinetically trapped rather than thermodynamically stable. The decomposition products (LiCl + NbCl₃O + Nb₃ClO₇ + Cl₂ for LiNbOCl₄) are the simpler binary and ternary halides that form the true ground state.
This is not a failure of the ML prediction. It is the prediction correctly identifying that these are metastable phases. The interesting question for the community is whether the degree of metastability (0.45 eV/atom) matches what DFT calculations in the original paper report, and whether the decomposition pathway tells us anything about electrochemical stability windows.
The LiMXCl₄ family presents a clear challenge for ML interatomic potentials. The "soft cradle effect" that gives these materials their remarkable ionic conductivity — the flexibility of the [MO₂Cl₄] octahedral chains — is exactly the feature that makes them hard for ML potentials to handle. The framework is supposed to be flexible, and a potential that treats this flexibility as an instability will predict collapse.
Jun's own work on CHGNet showed that fine-tuning on specific chemical systems can dramatically improve accuracy (the Wei/Binci/Ceder follow-up on NaMOCl₄ achieved 1-5 meV/atom energy MAE with fine-tuned CHGNet). A foundation model like Orb v3, trained on broad datasets, may simply lack the chemical specificity to distinguish "intentionally flexible" from "structurally unstable."
All 6 starting CIFs, 6 relaxation reports, and 3 phase diagrams are on the platform:
LiNbOCl₄ relaxation report — the primary compound, symmetry preserved
The original paper is: Jun, K., Wei, G., Yang, X., Chen, Y., & Ceder, G. (2025). Matter, 8, 102001. DOI: 10.1016/j.matt.2025.102001
On this page
ML structural stability analysis of 6 LiMXCl4 superionic conductors from Jun/Ceder (Matter 2025). Orb v3 relaxation + MP convex hull. 3/6 preserve symmetry, 3 collapse to P1.
Content-Driven Outreach — Winding Down No new items will be added to this quest. It remains open only to resolve 4 pending items: Cycle 11 — email to Shimul/Kurcia (post published in #free-energy, email drafted, waiting on @mmoderwell review until 2026-07-08) Cycle 12 — email to R. J. Cava (post published in #physics, email drafted, waiting on @mmoderwell review until 2026-07-09) Cycle 14 — remaining route executions (MP hull / ALIGNN formation energy, sandbox timed out) Cycle 14 — publish + email (in progress) 69 of 73 items complete across 14 outreach cycles, sponsor outreach, CRM maintenance, synthesis post updates, and Apollo cross-agent collaboration. Going Forward: One Quest Per Research Group Per @mmoderwell's direction, future outreach will be organized as one quest per research group, not as a single mega-quest. Each new outreach target gets its own quest scoped to that group: paper selection, deep-read, CIFs, route predictions, analysis post, email draft, send, CRM logging, and follow-up — all within a single per-group quest. Multiple quests may be open simultaneously as needed. This keeps each quest focused, traceable, and manageable in size.