cues-alex-nolemat-lowhull

CUES crystal generator — model card

The crystal-domain model behind the steering paper's cross-domain claims. Written to parity with the image setup (DiT-XL/2 + ResNet-50): architecture, training data, objective, release, and the exact metric definitions every cited number uses.

Model

Conditioning / I/O. A prompt is "<elements> | <natoms> | <bin-tags>", where the bin-tags condition the generation on coarse property bins:

• bg-{vlow,low,mid,high,vhigh} — band-gap bin
• hull-{vlow,low,mid,high} — energy-above-hull (stability) bin
• k-{…} — dielectric constant bin · rho-{…} — density bin

The decoder emits a structure encoding (anonymized formula + concrete formula, space group, and coordination/near-neighbor geometry) that is parsed to an ASE Atoms (lattice + sites). Naked prompts ("<elements> | <natoms> | ", no tags) are used for the distribution-steering experiments.

Training data

• Source: Alexandria-derived. The alex_nolemat_lowhull training set (alex_nolemat_lowhull_dataset_train.csv, ≈225 k structures), i.e. Alexandria structures, natoms < 20, restricted to a low energy-above-hull subset, with LeMat-Bulk overlap excluded (nolemat) so LeMat-Bulk can serve as the novelty reference at eval time.
• Each row: source, target, id, origin, label, e_above_hull, in_lemat — source is the prompt (elements | natoms | tags), target the structure encoding, origin = alex.

Objective

Supervised next-token cross-entropy: encoder-decoder translation from (composition + property-bin tags) to the structure encoding. No RL/preference objective in the base CUES model. (A separate aggressive DPO/IPO novelty fine-tune is used only for the Claim 4.4 off-support demonstration; it is not the CUES generator.)

Metric definitions (exact, as used in every cited number)

• stability — energy above hull (e_above_hull, eV/atom): a single MACE single-point energy evaluated against a convex hull whose Materials-Project reference entries are re-evaluated with the same MACE (self-consistent; cached per chemical system). stable = eah ≤ 0; metastable = 0 < eah ≤ 0.1. (chem.stability.compute_e_above_hull.) Needs MP_API_KEY + emmet-core. A 3-MLIP variant (orb _mpa + MACE + UMA, mean hull) is available for cross-checks; use orb _mpa not _omat.
• Combined = SUN + MSUN, over n_total (no validity gate — validity is a separate metric). SUN = stable ∧ novel ∧ unique; MSUN = metastable ∧ novel ∧ unique. Novelty (vs LeMat-Bulk) and uniqueness via lemat-genbench SUNMetric.
• BVS-GII (bond-valence self-consistency, valence units): the Global Instability Index — RMS over sites of (bond-valence sum from bond lengths, Brown–Altermatt calculate_bv_sum, minus the formal oxidation state). 0 = perfectly self-consistent; None for pure intermetallics (no anion).
• validity: SMACT charge-balance ∧ BVS-assignable ∧ minimum interatomic distance ≥ 0.5 Å (chem.validity.check_validity, one call).
• density (g/cm³): geometry only — total atomic mass / cell volume.

Reproducing every cited number

Scripts live in steering/crystals/ (mirrored from dielectric/eval/within_distribution_steering/):

• generate.py --target T + score.py — Combined steering (Claim 2).
• experiments/scan_extremes.py, natural_pools.py, method_compare.py — the steering range and the two target-hitting methods.
• scripts/claim1_survival_spectrum.py + claim1_survival_crystals.py — the Claim 1 carry-over ladder and density view.
• scripts/plot_pred_vs_obs.py, dpo_logprob_fig.py, make_figures.py — the paper figures.
• review1/ — the reviewer-requested CIs, convergence, joint baselines, and ladder prediction.

Reference data revisions (LeMat-Bulk, LeMat-Bulk-MLIP-Hull, UMA) are pinned in ARTIFACTS.md. The model is fetched from HF and hash-verified; nothing irreproducible is committed.

Scorers & statistics — pinned (camera-ready hygiene)

All numbers in the crystal track use these exact scorers/versions (full freeze in requirements.txt):

Sample sizes / seeds / CI method (consolidated):

• C-E1 density carry-over: n=30 chem, bootstrap-over-chemistries (B=10 000), 95% percentile CI.
• C-E1 DPO gap: n=250 structures, bootstrap, 95%.
• C-E1 Combined sweep: n=100/target, Wilson 95% on the proportion.
• C-E2 ladder: 24 chem × 3 seeds, M_data=15 / M_gen=30; bootstrap-over-chemistries (×seeds) 95%.
• C-E3 joint: N=120/condition × 3 seeds (→ 6 with the round-2 bump), Wilson 95% on pooled rate.
• C-E8 key sensitivity: re-analysis of the C-E2 per-item data, bootstrap-over-groups 95%.
• C-E9 ensemble: same 24 chem (seed 0), M_data=15 / M_gen=30, single-vs-ensemble r² on the same structures, bootstrap-over-chemistries 95%.
• C-E4 pre-registration: committed at git d96b827 before the C-E2 audit (08981f7); prior on a correct 3-rung ordering = 1/6.