Project 1 — Auditable formula-first screening of Li–S solid-electrolyte space

Project 1 turns 599 experimental ionic-conductivity records into a reproducible, composition-first ranking of a frozen Materials Project (MP) search space. The release is designed as an auditable statistical-screening study: it exposes split overlap, censoring, representation transfer, applicability domain, empirical risk heuristics, and the disposition of every queried MP entry.

The output is a ranking and evidence-state map, not a claim of ionic conductivity, novelty, synthesizability, or mechanism. No named "hero" candidate is promoted. MLIP-MD, new DFT calculations, synthesis, and wet experiments are explicitly outside Project 1 and belong in separate projects.

Study design

• OBELiX contributes 478 official training and 121 held-out records. Canonical, scale-invariant composition keys reveal two compositions shared across the official split. Strict held-out evaluation therefore removes the corresponding two training rows and uses 476 train / 121 test records; DOI overlap is zero.
• All 37 one-sided conductivity upper bounds (29 train, 8 test) are retained. The held-out point-MAE and one-sided censor-aware MAE are both 1.353 log units for this release.
• The primary representation contains 132 Magpie descriptors, number of elements, and Li atomic fraction (134 formula-only features). In matched five-fold composition-grouped CV, the 15-member perturbation ensemble reaches MAE 0.843 ± 0.136, RMSE 1.381 ± 0.176, and Spearman ρ 0.882 ± 0.012.
• On the strict held-out test, MAE is 1.353, RMSE 1.947, R2 0.411, Spearman ρ 0.660, and mean bias −0.363 log units. These errors preclude interpreting predictions as quantitative conductivity measurements. The predefined Li–S target-domain subset is smaller and weaker (24 records; MAE 1.585, Spearman ρ 0.375; bootstrap ρ interval −0.255 to 0.881).
• Matched strict-test baselines are reported rather than inferred from CV. The random forest and single base CatBoost have MAE 1.293 and 1.291, respectively, versus 1.353 for the perturbation ensemble. The paired cluster-bootstrap ensemble-minus-single-model difference is 0.062 (95% interval 0.020–0.106), so no strict-test ensemble advantage is claimed.
• Structural ablations are evaluated on identical folds. Raw cell conventions vary strongly across repeated compositions; even scale-invariant normalization leaves a residual outlier. Formula-only M0 is therefore the production representation because it is exactly computable for both experiments and MP.
• The frozen MP query uses database version 2026年04月13日 and returns 248 unique entries. Every entry receives exactly one disposition: 80 oxygen-scope exclusions, 19 redox/electronic-risk entries, 70 OBELiX-composition references, and 79 candidate entries.
• The 79 candidate structure entries collapse to a complete queue of 62 unique normalized compositions: 49 compositions have one MP structure, nine have two, and four have three. At the predefined q75 descriptive thresholds, 27 are extrapolative, 12 are high-score/lower-risk, and 23 are unflagged. These states describe evidence, not experimental or computational priority.
• Ensemble spread and applicability-domain (AD) distance are related empirical risk heuristics, not calibrated uncertainty. Their associations with error are reported for both the full strict test and the smaller Li–S target-domain subset; in the latter, all three correlations are weak and statistically inconclusive.

Five main figures are exported to figures/; one supplementary figure (boosting and learning curves) is exported to figures/supplementary/ under the same contract. Every file is a 900 dpi, RGB, pure-white, grid-free PNG with a 183 mm double-column width and height no greater than 170 mm. A shared Nature-informed visual system controls Arial-first typography, accessible colours, redundant marker shapes, evidence hierarchy, and panel spacing. No PDF, SVG, TIFF, or alpha channel is produced under the selected output contract; consequently, these files are visually prepared for manuscript review but are not a substitute for Nature's final editable-vector source. Stand-alone legends and panel-to-source mappings are in FIGURES.md.

1. Study design and accounting — measured evidence, split correction, and exhaustive 248-entry disposition.

   Figure 1

2. Model validation — target coverage, grouped CV, strict test, censoring, and cluster-bootstrap uncertainty.

   Figure 2

3. Transfer and risk — representation ablation, cell-convention audit, ensemble spread, and applicability domain.

   Figure 3

4. Global screening landscape — all 248 MP entries in descriptor, score, rank, and risk views without silent removal.

   Figure 4

5. Complete candidate atlas — all 62 compositions, their raw metrics, evidence states, and rank stability across all 15 ensemble members.

   Figure 5

OBELiX 599 records
 ├─ canonical composition + DOI + censor audit
 ├─ composition-grouped CV on official train
 └─ strict held-out test after two-row decontamination
 │
 ▼
formula-only 15-member perturbation ensemble
 ├─ seed + model-configuration variation
 └─ PCA / 5-nearest-neighbour applicability domain
 │
 ▼
frozen MP query: 248 entries
 ├─ 80 oxygen-scope exclusion
 ├─ 19 redox/electronic risk
 ├─ 70 OBELiX-composition reference
 └─ 79 candidate structure entries → 62 normalized compositions
 (49 ×ばつ 1 structure, 9 ×ばつ 2 structures, 4 ×ばつ 3 structures)

The MP chemical and thermodynamic filters define the population being ranked; they are not learned features. The production ensemble is trained on all 599 records only after evaluation is complete.

Python 3.11 or newer is required.

python -m venv .venv
source .venv/bin/activate
pip install -e '.[dev]'
p1screen evaluate
p1screen training-curves
p1screen screen
p1screen figures
p1screen verify --frozen

The wheel contains the Python workflow, not the writable research payload. For a non-editable installation, point it at an extracted Project 1 release before running data commands:

export P1SCREEN_ROOT=/path/to/project1_screening

To rebuild every derived artifact from the vendored OBELiX splits and frozen MP snapshot:

p1screen rebuild

Refreshing MP is an explicit, reviewed operation because database content can change:

export MP_API_KEY='...'
p1screen refresh-mp
p1screen rebuild

The refresh command accepts the key only through MP_API_KEY, requires the reviewed database version and 248-entry count, and stores scalar metadata only. No crystal structures or credentials are distributed.

The numbered Python files remain thin compatibility wrappers around the same CLI. p1screen is the authoritative interface.

src/p1screen/ package: data, features, models, screen, figures, gate
data/train.csv vendored OBELiX official training split
data/test.csv vendored OBELiX official test split
data/mp_snapshot.csv frozen 248-entry minimal MP metadata snapshot
data/metrics.json evaluation results and audit statistics
data/model_manifest.json 15 member definitions and SHA-256 checksums
data/release_manifest.json reviewed offline release checksums
artifacts/models/ 15 serialized production ensemble members
mp_ledger.csv exhaustive entry-level 248-row ledger
candidate_queue.csv complete 62-composition candidate table
source_data/ panel-ready quantitative CSVs
figures/ five main 900 dpi RGB PNG files
figures/supplementary/ Supplementary Fig. S1 (boosting and learning curves)
tests/ unit and frozen-contract tests

The offline gate rejects stale figure/source inventories, missing candidates, checksum mismatches, non-RGB images, non-white canvases, wrong DPI, database drift, incomplete dispositions, and credential-like tracked files.

• OBELiX is small and heterogeneous; measurement temperature, processing, microstructure, disorder, and interface effects are not uniformly represented.
• A formula-only model cannot resolve polymorphs, defects, phase mixtures, electronic leakage, or kinetic stability.
• The held-out MAE is approximately 1.35 orders of magnitude. The MP output must be used for population ranking, not conductivity prediction in physical units.
• In the 24-record target-domain subset, MAE rises to approximately 1.58 and R2 is negative; its broad bootstrap intervals make transfer conclusions tentative.
• Ensemble spread and AD distance identify relative risk but are not calibrated confidence intervals and cannot guarantee error out of domain.
• The oxygen and redox/electronic rules are explicit scope filters, not universal chemical impossibility statements.
• An ICSD identifier is a provenance cue only. No novelty claim is made.
• No downstream DFT, molecular dynamics, synthesis, or wet-experiment outcome is used to evaluate or retroactively tune Project 1.

OBELiX data are credited to Therrien et al., Digital Discovery (2026), doi:10.1039/D5DD00441A, and retain their CC BY 4.0 terms. Materials Project metadata must be cited using the references listed in THIRD_PARTY_NOTICES.md. Repository code is MIT licensed; data and third-party metadata are not relicensed by that grant.

See METHODS.md, DATA_DICTIONARY.md, REPRODUCIBILITY.md, and CITATION.cff.