Current project snapshot — 2026-11-02

The assembled dataset contains 10,676 distinct value/context points, 13,379 accepted evidence observations, and 37 properties. 5,121 contexts have numeric temperatures; 5,555 lack supported numeric conditions. Of 7,950 measured separator/component contexts, 2,571 have numeric temperatures. Bulk references remain separately identified.

This pass added 25 MD-reference contexts and 40 temperature-resolved separator observations (including 32 calibrated figure estimates). Source review corrected 105 observations and held 65 ambiguous observations for review. All withdrawn evidence remains traceable. The total temperature count fell after removing incorrect inherited conditions.

Run .venv/bin/python scripts/assemble_dataset.py --check-replay to rebuild from the existing local source archive. Validation: 167 tests pass; 42 generated artifacts replay byte-for-byte. A fresh clone requires the source archive for a full rebuild. Arbitrary PDF/figure interpretation still requires reviewed calibration and attribution.

See the comprehensive project guide, dataset, and figures. Earlier dated sections below are historical snapshots.

Historical thermal collection and temperature recovery — 2026-10-30

Current local exports: 10,582 distinct contexts, 13,258 evidence observations, 5,110 with numeric temperature (+2,419). Added 2,200 FSRI bulk-reference contexts and 23 paper contexts; recovered conditions for 196 existing contexts. Full data remain in datasets/unique_points.csv; temperature_resolved_points.csv is the explicit-temperature subset. 5,472 still lack numeric test conditions, including 676 temperature-response properties. Bulk references and separator performance remain distinguished. See the current report and datasets/temperature_gap_audit.csv for evidence, limits and next actions. Earlier sections below are historical.

Earlier collection snapshot — 2026-10-26

Current local exports contain 8,320 distinct contexts and 10,996 accepted observations, using separator-v32 / schema 2. The pass admitted 2,338 new measurement contexts; the 3,000-additional-point target was not met. See the latest collection report for material and temperature/property coverage, PDF attribution and validation. Earlier dated sections below describe earlier snapshots.

ChE126 battery separator properties

Public dataset: https://huggingface.co/datasets/archiitecture/che126

Code and research notes: https://github.com/archiitectures/che126_data_science

Earlier independent review — 2026-10-23

That reviewed snapshot contained 8,977 accepted observations and 5,982 distinct scientific value/context groups, across 23 properties, after the independent review (separator-v24, schema 2). The review moved 15 relative/dispersion values out of accepted absolute-property records and corrected context grouping. More context groups do not mean more measurements were collected.

Use master_observations.csv/.parquet for all accepted evidence and unique_points.csv/.parquet for the grouped view. These are tidy tables, one property per row. unique_point_evidence.json preserves all supporting observation IDs. Method, measurement context, shrinkage basis, material grade, replicate/curve identity and uncertainty now participate in grouping. Equal recorded contexts can still represent separate experiments; patent families have not been fully linked.

The source citations include 275 patent publication URLs, 13 literature URLs, 28 Celgard datasheets and 40 FSRI file URLs. There are 349 import/source files: one curated file can contain several primary sources. These are distinct counting units, and patent publication count is not patent-family count.

See independent review and audit results. The local exports have not been republished to Hugging Face as part of this review. Missing values remain unknown; bulk references and calculated/specification records must be filtered appropriately. For structured imports, original_value/original_unit describe the imported CSV cell; primary-document values and hashes can additionally reside in notes and dated curation evidence. A CSV hash alone does not verify the original paper.

The legacy dry/wet tables and their 35-column dictionary below are a separate older format, not the schema of the master/unique tables. Earlier reports are snapshots.

Current material identity audit — 2026-10-23

The stored sources have been reprocessed to link separator material recipes to individual samples, including explicit recipe inheritance, layered composition tables and reviewed source passages. Current counts and remaining gaps are in the material identity report. Earlier collection counts below describe the snapshot before this identity reconciliation.

Use unique_points.csv for distinct normalized property/context points and master_observations.csv for the full accepted source observations. Corrected material labels can split or merge distinct contexts; they do not represent newly collected measurements. The reconciliation checks preserve original values, units, temperatures, sample labels, directions and protocols.

The CSVs add material_identity_status, material_identity_evidence and material_identity_reason. A missing polymer is accompanied by a specific linking/evidence gap; it does not by itself establish that the source omitted its material. The per-point identity audit and the report's observation lineage make these assignments traceable. Polymer families outside PE/PP, including identified fiber composites, remain explicit categories. Current figures and their reproduction code live in visualizations.

Historical collection snapshot — 2026-10-23 balanced pass

The completed bounded pass contains 3,961 distinct normalized value/context points from 8,992 accepted source observations, net +1,802 contexts from the 2,159-point baseline. Validation confirms all baseline contexts were preserved, complete evidence links, CSV/Parquet round-trip consistency, and byte-equal dataset exports. The 2,000-additional-context target was not met. The increase comprises 587 contexts recovered from cached sources/reviewed cached imports and 1,215 from new sources. Distinct contexts are not independent experiments.

There are 23 property categories and 998 separator points with numeric measurement/exposure temperatures, up from 488. Across separator records, 20 setpoints span 23–300 °C. Identified PE, PP, PE/PP, and PP/PE/PP records cover 14 setpoints from 23–180 °C; the 250 and 300 °C records have unresolved polymer identity and do not demonstrate high-temperature performance for identified polyolefin types. Temperature marginals improved, but occupied cells in the fixed baseline property × temperature × role grid increased only from 19 to 20. Full cross-property temperature coverage remains sparse.

The parser version for this pass is separator-v19. Property categories are values in the tidy property field, not literal wide columns. New source formats include patent HTML tables, primary Europe PMC XML, and a reviewed primary author-hosted PDF. One new specific-heat-capacity category has a single observation. One separator volumetric heat-capacity value is explicitly calculated as density × specific heat; it is not an independent experiment. An unnumbered “room temperature” is not assigned a numeric temperature, and no interpolation is performed.

See the balanced collection report, run report, and run README for property and temperature growth tables, methodology, and limitations. The 2026-10-19 expansion report and earlier reports remain available as historical snapshots. Pending review records are excluded; source restrictions remain recorded and no dataset-wide license is asserted.

Separate dry and wet datasets

• dry_separators.csv / dry subset: measurements without battery electrolyte soaking, including explicitly labeled bulk-sheet reference rows. Leave electrolyte and soaking fields blank. Select sample_form != bulk_sheet_reference to exclude these reference observations from separator-specific analysis.
• wet_separators.csv / wet subset: separators measured while soaked in battery electrolyte. Record composition and soaking conditions whenever reported.

Dry/wet here describes the specimen's state during measurement, not whether the separator was manufactured using a dry or wet process. Do not infer state from manufacturing process. Sources with unclear measurement state need clarification before entering either table. If a soaked specimen is dried before testing, put that measurement in dry and describe its history in notes.

Both CSVs share the same 35 columns (the original 27 plus eight thermal/reference fields). Pair dry and wet measurements from the same material or specimen using a stable sample_id; every row has its own record_id. Hugging Face configurations keep the subsets separate. A library's automatically assigned train split is a storage convention, not an agreed evaluation split.

Row and missing-value conventions

One row describes a source-reported observation under a specific measurement condition and direction. Multiple properties may share a row only when source, specimen, temperature, direction, and protocol apply to all those values. Use separate rows for different temperatures, directions, tests, or replicates; leave unrelated property cells blank rather than copying values into new tests.

An empty cell means unreported or not applicable, never zero. Store numeric values without unit strings. porosity_pct uses 0–100 percent, not a 0–1 fraction. Do not guess room temperature. Preserve original values/units, conversion notes, uncertainties, and unusual protocols in notes until dedicated fields are needed. Only enter strength in a canonical column when its definition and unit match; for example, tensile force in N is not tensile stress in MPa.

Column dictionary

Column	Meaning
record_id	Unique observation identifier across both files.
sample_id	Stable specimen/material identifier for linking conditions.
material	Polymer/material composition, such as PE or PP.
separator_type	Construction, multilayer arrangement, or coating.
product_name	Manufacturer/product designation when known.
temperature_C	Temperature during the reported property measurement, °C.
conditioning_temperature_C	Prior heat-treatment temperature, °C, if distinct.
conditioning_duration_h	Duration of that prior heat treatment, hours.
porosity_pct	Reported void fraction, percent; retain method/context.
tensile_strength_MPa	Ultimate tensile stress, MPa; exclude yield stress.
puncture_strength_N	Force at puncture, N; exclude pressure or stress.
thickness_um	Separator thickness, µm, under the recorded condition.
pore_size_um	Reported pore size, µm; identify mean/median/max in notes.
measurement_direction	MD, TD, through-plane, or source-specific direction.
porosity_method	Porosity measurement method and relevant conditions.
strength_method	Strength test method, geometry, rate, and loading details.
size_method	Thickness/pore-size method, load, and size statistic.
electrolyte_composition	Salt, concentration, solvents/ratios, and additives.
soaking_time_h	Electrolyte soaking duration, hours; wet observations only.
soaking_temperature_C	Electrolyte soaking temperature, °C; wet only.
source_id	Stable source identifier, shared by observations from that source.
source_type	Paper, datasheet, patent, simulation, or public_materials_database; identify origin explicitly.
source_url	DOI link or stable source/document URL.
source_locator	Document version and page, table, figure, or simulation run.
value_type	Measured, typical, minimum, maximum, digitized, simulated, or measured_and_calculated; notes identify calculated fields.
replicate_id	Reported replicate identifier; blank for unspecified replicates.
notes	Original units/values, uncertainty, thermal history, and caveats.
sample_form	Explicit specimen category; bulk_sheet_reference for these FSRI observations.
source_condition	Source's original conditioning terminology and its meaning.
reference_thickness_um	Instrument file-header specimen thickness, µm; temperature unspecified. It is not a thermal thickness curve.
thermal_conductivity_W_m_K	Reported HFM Average Cond, W/(m K).
volumetric_heat_capacity_J_m3_K	Measured HFM heat capacity per unit volume, J/(m³ K).
specific_heat_capacity_J_kg_K	Calculated mass-specific heat capacity, J/(kg K); retain conversion in notes.
reference_density_kg_m3	Arithmetic mean of the source's three density replicates used for conversion, kg/m³; not a temperature-dependent density measurement.
thermal_method	Thermal-property apparatus, test, and relevant protocol.

Temperature during measurement, prior thermal conditioning, and soaking temperature are distinct quantities. A datasheet thickness or porosity should not inherit a temperature from an unrelated shrinkage test.

Curation and access

Hugging Face is the preferred home for the growing property tables. The code repository keeps small versioned tables, this dataset card, and acquisition utilities. The FSRI result-table excerpts are pinned to a source commit for reproducibility. Future records must retain provenance and distinguish literature, manufacturer, digitized, and simulation-derived values.

Once these files are published on the Hub, use the dry and wet configuration names with datasets.load_dataset. The wet subset is empty. FSRI bulk-sheet references alone do not establish separator behavior.

No dataset-wide license is declared yet. Source-specific reuse terms must be recorded and reviewed as real records are curated; public hosting does not grant a new license to third-party source documents.

FSRI polyethylene coverage and provenance

Source: UL Research Institutes, Fire Safety Research Institute, Materials and Products (MaP) Database. Website, experimental methods, and repository citation. Source commit: a432697e4b7d40f642b5cc8f384506439ebb32df.

Material	Conductivity observations	Heat-capacity observations	Total
HDPE	18	12	30
LDPE	18	12	30

All numeric rows in the 12 HFM SI result tables are included: conductivity at 15.01 and 45.01 °C, and heat capacity at 10.00, 20.01, 30.01 and 40.01 °C. Reported temperatures are not rounded to nominal setpoints. Each conductivity file has six results, alternating between its two temperatures three times. These repeated setpoints are retained separately and must not be counted as six independent specimen replicates. Each heat-capacity file has four reported results; its five programmed setpoints include an initial step without a reported heat-capacity result, so no fifth value is invented.

Coverage is the HFM thermal-property series. STA/MCC powder decomposition and combustion time series are not imported as separator measurements. Each row links to its exact source file at the pinned commit, records its blob SHA and SI result-table data-row number, and identifies the source replicate.

FSRI calls these samples “Wet” because they were unconditioned with ambient moisture. They are included in our dry table as bulk-sheet references without battery electrolyte, by project decision. The wet CSV is reserved for actual battery-electrolyte-soaked measurements. This does not imply that FSRI performed oven drying on these specimens.

Instrument header thicknesses (3170–3380 µm) are recorded only in reference_thickness_um. The source does not establish a temperature-dependent thickness measurement, so thickness_um stays blank along with separator porosity, strength, and pore size.

The raw heat-capacity unit is J/(m³ K). To provide J/(kg K), divide by the mean of the three source density records: HDPE (974 + 970 + 970) / 3 and LDPE (925 + 930 + 927) / 3 kg/m³. This follows the mean-density normalization in FSRI's HFM processing script. Calculated specific heat is rounded to 0.001 J/(kg K); this rounding does not represent an uncertainty estimate. The original volumetric measurement and density inputs are preserved. value_type=measured_and_calculated identifies these rows. LDPE density file column labels say HDPE_WET despite their LDPE paths; the inconsistency is recorded in affected row notes.

The source excerpts retain complete SI result tables, file-header excerpts, and all six density CSVs. che126_data_science.data_mining.fsri.polyethylene_reference_rows(snapshot) reproduces the reference records without network access or third-party runtime dependencies. Original UTF-16 file blob hashes were verified during acquisition.

Polymer subtype metadata — 2026-10-30

The master, unique-point and temperature-resolved tables now include polymer subtype, scoped morphology, assignment status/evidence, and unverified source candidates. See the probe report for counts, limitations and the required postprocessing command after rebuilding exports. Polymer grade and finished membrane crystallinity are separate attributes.

Stored-source property and morphology recovery — 2026-10-30

Current exports contain 10,654 unique contexts from 13,371 observations. This pass added expansion strains, compression/recovery, and method-specific crystallinity measurements. See the source probe report for counts, patent morphology coverage, coefficient declaration flags and the remaining gap from amorphous PE simulation. polymer_crystallinity_measurements preserves sample/component scope and multiple assay methods as a JSON list.