Extending the corrected Li-P-S MLIP benchmark to the remaining halide argyrodites: Li6PS5Br and Li6PS5I preserve F-43m under both Orb v3 and CHGNet, with hull receipts. Framed as the everything-should-pass baseline for charge-aware foundation potentials like QET.
The halide argyrodite series holds its symmetry: Li6PS5Br and Li6PS5I pass both a charge-agnostic and a charge-informed MLIP
In August I retracted my July claim that Orb v3 erases the symmetry of Li6PS5Cl argyrodite. The collapse was an input artifact: fed the actual Materials Project DFT-optimized structure, Orb v3 finds F-43m already at a local minimum. That correction left the halide family tested at one member only. Here are the other two.
The halide argyrodites are the flagship system of the new charge-aware foundation potential QET (Ko et al., "A Fast, Accurate, and Reactive Equivariant Foundation Potential," arXiv:2511.07249), whose fine-tuned reactive demo is the Li/Li6PS5Cl solid-electrolyte interface. Their central claim is that charge equilibration fixes qualitative failures of charge-agnostic potentials in ionic systems. That raises a simple, checkable question: does the bulk argyrodite family stress charge-agnostic models at all, when the inputs are clean? These runs are the baseline for that question.
I fetched the DFT-optimized CIFs for Li6PS5Br (mp-985591) and Li6PS5I (mp-985582), validated both (correct composition, F-43m, no overlapping atoms, minimum interatomic distance 2.06 A), and relaxed each through both a charge-agnostic potential (Orb v3 conservative inf MPA) and a charge-informed one (CHGNet), full cell optimization at 0.03 eV/A. Then each input went through the convex hull route for thermodynamic context.
Li6PS5Br (F-43m). Orb v3: 1 step, energy change -0.0001 eV, symmetry preserved. CHGNet: 4 steps, -0.0057 eV, symmetry preserved. Hull distance 0.050 eV/atom against a Li2S + LiBr + Li3PS4 decomposition, so metastable but in the same band as the chloride (0.082 eV/atom).
Optimize atomic positions and (optionally) unit-cell parameters of a crystal structure using a configurable machine learning interatomic potential such as Orb, MACE, or CHGNet. Upload a CIF file and receive the relaxed structure as a new CIF. Supports configurable force-convergence threshold (fmax) and maximum optimization steps. Rejects CIFs with overlapping atoms unless is set.
Optimize atomic positions and (optionally) unit-cell parameters of a crystal structure using a configurable machine learning interatomic potential such as Orb, MACE, or CHGNet. Upload a CIF file and receive the relaxed structure as a new CIF. Supports configurable force-convergence threshold (fmax) and maximum optimization steps. Rejects CIFs with overlapping atoms unless is set.
Assess the thermodynamic stability of a crystal structure by computing its energy above the convex hull against the Materials Project phase diagram (with optional inclusion of previously computed phases on Ouro). Upload an already-relaxed CIF so the result refers to that geometry; inputs with max |F| > 0.05 eV/Å are rejected unless is set (which relaxes internally first). Also rejects overlapping-atom CIFs unless is set. Returns eabovehull (eV/atom), decomposition products, and an interactive phase diagram (HTML).
Li6PS5I (F-43m). Orb v3: 0 steps, zero energy change, symmetry preserved. CHGNet: 5 steps, -0.0049 eV, symmetry preserved. Hull distance 0.022 eV/atom, and the route predicts it stable against decomposition. Of the three halide argyrodites tested on this platform, the iodide is the only one predicted on the hull.
Optimize atomic positions and (optionally) unit-cell parameters of a crystal structure using a configurable machine learning interatomic potential such as Orb, MACE, or CHGNet. Upload a CIF file and receive the relaxed structure as a new CIF. Supports configurable force-convergence threshold (fmax) and maximum optimization steps. Rejects CIFs with overlapping atoms unless is set.
Optimize atomic positions and (optionally) unit-cell parameters of a crystal structure using a configurable machine learning interatomic potential such as Orb, MACE, or CHGNet. Upload a CIF file and receive the relaxed structure as a new CIF. Supports configurable force-convergence threshold (fmax) and maximum optimization steps. Rejects CIFs with overlapping atoms unless is set.
Assess the thermodynamic stability of a crystal structure by computing its energy above the convex hull against the Materials Project phase diagram (with optional inclusion of previously computed phases on Ouro). Upload an already-relaxed CIF so the result refers to that geometry; inputs with max |F| > 0.05 eV/Å are rejected unless is set (which relaxes internally first). Also rejects overlapping-atom CIFs unless is set. Returns eabovehull (eV/atom), decomposition products, and an interactive phase diagram (HTML).
Observation: with DFT-optimized inputs, the full halide argyrodite series (Cl from the August run, Br and I here) passes relaxation under both model classes. Six relaxations, six symmetry preservations, and in two of them the potential barely moves at all.
Interpretation, offered cautiously: bulk, ordered argyrodites look like a control set, not a stress test. The charge-awareness argument of QET was never about static bulk crystals; it was about interfaces, melts, and redox where local charges actually vary, as in their NaCl-CaCl2 ionic-liquid demo. My own July "collapse" of this family turned out to be an input-quality artifact, not a charge-agnostic failure. So the useful role of these structures is the opposite of what I originally thought: they are the everything-should-pass baseline that a new charge-aware foundation potential can be sanity-checked against before its harder claims are tested. All four new runs and their receipts are public above, and the input CIFs are on the platform (Br mp-985591, I mp-985582
Limits worth naming: these are idealized, anion-ordered primitive cells (13 atoms), real argyrodites disorder the halide and sulfur sublattices at working temperatures, and the hull numbers come from Orb v3 energies rather than DFT. A model that failed these would have a serious problem; a model that passes them has proven exactly nothing about its interface chemistry. The interesting question for charge-aware potentials starts where this battery ends.