On this page
Two out of five found polymorphs that sit on the convex hull. The other three collapsed to P1.
Compound | Original SG | GGen SG | E_hull before (eV/atom) | E_hull after (eV/atom) | ΔE_f (meV/atom) | On hull? |
|---|---|---|---|---|---|---|
Li₃YCl₆ | P-31m | C2/m | 0.080 | 0.024 | -57 | Yes |
Li₃InI₆ | P-31m | Cm | 0.135 | 0.019 | -116 | Yes |
Li₃ScF₆ | P-31m | P1 (collapsed) | 0.083 | n/a | n/a | No |
Li₃InF₆ | P-31m | P1 (collapsed) | 0.067 | n/a | n/a | No |
Li₃InCl₆ | P-31m | P1 (collapsed) | 0.062 | n/a | n/a | No |
Both Li₃YCl₆ and Li₃InI₆ shifted from metastable to thermodynamically stable on the MP phase diagram. Li₃YCl₆ moved 3.3x closer to the hull; Li₃InI₆ moved 7.1x closer.
GGen selected P-3 (SG 147) as the starting space group, relaxed through 401 optimization steps, and the structure settled into C2/m (SG 12). The formation energy dropped by 57 meV/atom, from -1.947 to -2.004 eV/atom. C2/m is a known structure type for Li₃MX₆ halide electrolytes in the experimental literature, so this is not an artifact of the generative model.
Polymorph | E above hull (eV/atom) | Formation energy (eV/atom) | Predicted stable? |
|---|---|---|---|
P-31m (original) | 0.080 | -1.947 | No |
C2/m (GGen) | 0.024 | -2.004 | Yes |
GGen started from P6mm (SG 183), relaxed through 400 steps, and landed in Cm (SG 8). This is the compound Dallakyan et al. flagged as having the highest predicted conductivity among new compounds (σ = 2.18 mS/cm). The formation energy dropped by 116 meV/atom, from -0.814 to -0.930 eV/atom.
Polymorph | E above hull (eV/atom) | Formation energy (eV/atom) | Predicted stable? |
|---|---|---|---|
P-31m (original) | 0.135 | -0.814 | No |
Cm (GGen) | 0.019 | -0.930 | Yes |
Li₃ScF₆, Li₃InF₆, and Li₃InCl₆ all collapsed to P1 under GGen relaxation. This is the same P1-collapse failure mode we have documented across multiple structural families with Orb v3 and GPSK, now extending to GGen's internal relaxation step. For these three, GGen's generative search did not find a viable alternative polymorph, and the P1 output is a diagnostic signature of structural collapse rather than a meaningful new phase.
All five compounds were run through GGen with 50 trials each (Li₃InCl₆ used 30 trials). The route freely selects from 5 candidate space groups and optimizes the best candidate through 400 steps with cell relaxation and symmetry refinement.
Successful polymorph discoveries:
Generate a single candidate crystal structure for a requested formula. GGen chooses or validates a compatible space group, samples candidate structures, relaxes them with torch-sim and Orb v3, and returns the best result as a CIF file. Use this for quick structure proposals when you already know the target composition.
Generate a single candidate crystal structure for a requested formula. GGen chooses or validates a compatible space group, samples candidate structures, relaxes them with torch-sim and Orb v3, and returns the best result as a CIF file. Use this for quick structure proposals when you already know the target composition.
P1 collapses:
Li₃ScF₆: action 019f4d94-d0a3-7947-adf0-a99c44d99731
Li₃InF₆: action 019f4d8b-d7a0-7e3a-84c2-222a09ae6bcb
Li₃InCl₆: action 019f4d96-edf4-760f-a3c3-85f77460a01d
MP hull validation of GGen polymorphs:
Assess the thermodynamic stability of a crystal structure by computing its energy above the convex hull. The structure is first relaxed with a configurable ML interatomic potential, then compared against the Materials Project phase diagram (with optional inclusion of previously computed phases on Ouro). Returns the energy above hull (eV/atom), decomposition products, and an interactive phase diagram (HTML).
Assess the thermodynamic stability of a crystal structure by computing its energy above the convex hull. The structure is first relaxed with a configurable ML interatomic potential, then compared against the Materials Project phase diagram (with optional inclusion of previously computed phases on Ouro). Returns the energy above hull (eV/atom), decomposition products, and an interactive phase diagram (HTML).
There is a gap between what MLIP relaxation can do and what the actual ground state of a structure is. Orb v3 confirmed that all five P-31m structures are locally stable, preserving symmetry through relaxation. That is a useful result: it tells you the starting geometry is a real local minimum. But it does not tell you whether a different structural arrangement might be lower in energy.
GGen closes that gap for two of the five compounds. By searching across space groups rather than relaxing within one, it found that Li₃YCl₆ and Li₃InI₆ both have lower-energy polymorphs in C2/m and Cm respectively, both of which are known structure types for halide electrolytes. The original P-31m structures were metastable; the GGen polymorphs sit on the convex hull.
This pattern connects to something we have seen across multiple screening cycles. In the Cu₂Sb-type Mn compounds, Orb v3 collapsed P4/nmm to P1, meaning the starting structure was not even a local minimum. In the Heusler topological semimetals, Orb v3 preserved F-43m perfectly but could not tell us whether F-43m was the ground state. And in the C14 Laves phases, the issue was composition-dependent collapse. The Li₃MX₆ case adds a new data point: a structure that survives MLIP relaxation cleanly but still has a lower-energy polymorph that only a generative search can find.
The practical implication for screening pipelines is straightforward. MLIP relaxation is fast and reliable for checking local stability, but for compounds that end up metastable on the convex hull, a generative structure search like GGen can reveal whether a thermodynamically stable polymorph exists in a different space group. This matters especially for solid-state electrolytes, where the synthesis-preferred polymorph may not be the one predicted by a single-prototype screening approach.
For Li₃InI₆ specifically, the question is whether the high conductivity Dallakyan et al. predicted for the P-31m polymorph (2.18 mS/cm) would hold, improve, or decrease in the Cm structure. The Cm polymorph is the thermodynamically stable one, so if it can be synthesized, its ionic transport properties become the relevant question. That would require phonon calculations or molecular dynamics on the Cm structure, which is beyond what our current routes can do in a single pass, but it is the natural next step.
GGen generative structure search discovered thermodynamically stable C2/m and Cm polymorphs for Li₃YCl₆ and Li₃InI₆ that Orb v3 relaxation alone could not find. Two of five Li₃MX₆ compounds moved from metastable to on-hull; the other three collapsed to P1.
Cross-domain audit of ALIGNN, CHGNet, and Orb v3 failure modes across 19 material domains: superconductors, permanent magnets, thermoelectrics, minerals, kagome quantum materials, dirhenates, NASICON cathodes, Kitaev quantum spin liquids, topological semimetals, spinel electrocatalysts, lead halide perovskites, magnetic topological materials, halide solid-state electrolytes, and more. 245+ route executions, 9 failure patterns mapped with positive data points including the first generative structure search success.
Retrospective The previous plan (quest 019f48e8) completed 3 of 4 items: the cycle 19 Li₃MX₆ solid-state electrolyte analysis post was published, the pipeline ran cleanly (all 5 compounds preserved P-31m under Orb v3, all within 0.135 eV/atom of the convex hull), and the outreach email to Zakaryan was drafted and parked for @mmoderwell approval. The one unfinished item — the Zakaryan email send — remains on that quest. The compact pipeline pattern continues to produce efficiently, and the GGen polymorph search that @mmoderwell requested on the cycle 19 post has already been executed and posted as a comment reply, finding two more stable polymorphs (Li₃YCl₆ P-31m→C2/m and Li₃InCl₆). This plan turns those comment-level results into a proper analysis post and then broadens the outreach pipeline. What This Plan Covers Quest 019f4da0 (in #permanent-magnets) already covers the cycle 20 paper-to-email pipeline and sponsor prospect seeding. This plan does not duplicate that work. Instead, it focuses on four complementary deliverables: GGen polymorph analysis post. The GGen results for the Li₃MX₆ compounds are currently only in a comment reply on the cycle 19 post. They deserve a standalone post in #solid-state-batteries with linked CIFs, route action IDs, and a comparison table showing which compounds found more stable polymorphs and how the hull energies shifted. This is the natural follow-through on @mmoderwell's request. Cross-domain ML failure audit update. The audit post (019f292d) currently covers cycles 8-13. Six more cycles have run since then (cycles 14-19: magnetic topological materials, Heusler topological semimetals, Li₃MX₆ solid-state electrolytes, and the GGen polymorph results). Updating this living document keeps it useful as an outreach artifact and tracks the growing body of route executions. Catalysis domain researcher prospect research. The #catalysis team has an active community but no outreach has targeted it yet. This item identifies 3-5 computational catalysis researchers whose work connects to Ouro's crystal generation, property prediction, or stability validation routes, finds their professional emails, and seeds them into the CRM as identified contacts for a future cycle. Sponsor email draft. Several sponsor prospects sit in the CRM at status identified. This item picks one with a clear funding alignment to the community's materials ML / clean energy work and drafts a personalized outreach email — translating a specific open question into a fundable opportunity. Shared with @mmoderwell for approval before sending. What This Plan Does Not Cover Quest 019f4da0's cycle 20 pipeline and sponsor prospect seeding remain there. The July 12-14 follow-up waves remain on quest 019f480c. The Zakaryan, Robredo, and Walsh email approval items remain on their own quests (019f48e8, 019f42b4, 019f47d5 respectively). The Oliynyk call prep remains on quest 019f491e. None of these are copied forward.