fishmoon1234/Polymer_Fire_Properties

Polymer Fire Properties

Repeat-unit polymer structures written as pSMILES, paired with four fire properties:

Symbol	Property	Unit
tig	time to ignition	s
phrr / pkhrr	peak heat release rate	kW m⁻²
SEA	specific extinction area	m² kg⁻¹
CO	carbon monoxide yield	kg kg⁻¹

Three datasets are provided: a large corpus of enumerated repeat-unit structures (v1), a simulated corpus labelled with the Fire Dynamics Simulator (v2), and a small set of bench-scale cone-calorimeter measurements collected from the literature, together with the ten cross-validation folds used to evaluate property prediction on it.

Structures are canonical SMILES with each open end of the repeat unit written as a bare asterisk, for example *CC(*)C for polypropylene.

Files

Path	Rows	Contents
v1_1.7M_synthetic_structure.smi	1,735,755	Enumerated repeat-unit structures, unlabelled
v2_110k_FDS.smi	109,993	Simulated cone-calorimeter cases
experimental_flux_thickness.smi	203	Measured cone-calorimeter data, 49 polymers
10folds_data/	—	Ten-fold splits of the experimental data, per property

---

v1_1.7M_synthetic_structure.smi

One pSMILES string per line, no header and no other columns.

The corpus contains 1,732,431 distinct strings over 1,735,755 lines: 1,077 strings occur more than once, for 3,324 duplicate lines in total. Deduplicate if your use requires unique structures.

v2_110k_FDS.smi

Whitespace-separated, no header, nine columns:

#	Column	Unit	Notes
1	index	—	row counter
2	name	—	POL_#####, one identifier per repeat unit
3	psmiles	—	canonical, bare *
4	thickness	mm	1.0 – 25.0
5	flux	kW m⁻²	20.0 – 100.0
6	tig	s	see the note on non-ignition below
7	pkhrr	kW m⁻²
8	sea	m² kg⁻¹
9	co	kg kg⁻¹

55,000 distinct repeat units, each simulated at two exposures: one incident flux drawn from 20–50 kW m⁻² and one from 50–100 kW m⁻², each at a specimen thickness drawn from 1–25 mm. Simulations use the solid-phase-only pyrolysis model of the Fire Dynamics Simulator, version 6.10.1. Ignition is taken at a critical heat release rate per unit area of 5 kW m⁻². From each simulated curve, tig is the time at which the heat release rate per unit area first exceeds 5 kW m⁻², and pkhrr is the maximum of that curve.

Two things to handle before training on this file.

Non-ignition is censored, not measured. In 1,566 rows the specimen never ignited within the simulated horizon, and tig carries the horizon value of exactly 900.0 s rather than a time to ignition. These rows are identified by tig == 900.0: the longest ignition time among cases that did ignite is 899.0 s, so the test separates them exactly. Treat them as right-censored at 900 s, or drop them. Fitting a regressor to them as if 900 s were an observation will bias the long-ignition end.

SEA and CO do not vary with the exposure. They are not inputs to the pyrolysis model; they are assigned once per repeat unit from the structure–property relationships. Both rows of a given polymer therefore repeat the same sea and co, and a model that conditions them on thickness or flux is fitting noise.

experimental_flux_thickness.smi

Space-separated with a header row:

smiles_canonicalized thickness flux tig pkhrr sea co

203 measurements over 49 polymers, collected from the literature, at specimen thicknesses of 1.6–25.0 mm and incident fluxes of 20–100 kW m⁻².

Coverage is uneven, and empty cells read as NaN:

Property	Rows measured	Polymers measured
tig	203	49
pkhrr	203	49
sea	157	44
co	96	32

This file holds one row per (polymer, thickness, flux). Where a source reports replicate measurements at the same condition they are collapsed here; the fold files below keep the replicates, which is why they hold more rows.

10folds_data/

Ten-fold splits of the experimental data, one folder per property, each with data_0 … data_9 holding train.csv and test.csv.

Folder	Columns	Rows per fold (train / test)	Distinct polymers
tig_exp_data/	smiles_names, smiles_canonicalized, thickness, flux, tig	206 / 22	49
phrr_exp_data/	smiles_names, smiles_canonicalized, thickness, flux, pkhrr	206 / 22	49
SEA_exp_data_avg/	smiles_canonicalized, SEA	40 / 4	44
CO_exp_data_avg/	smiles_canonicalized, CO	29 / 3	32

Note that the target column inside phrr_exp_data/ is named pkhrr, not phrr.

tig and pkhrr are modelled per measurement, so their folds carry the exposure conditions and retain the replicate measurements: 228 rows in all, against the 203 rows of experimental_flux_thickness.smi. SEA and CO are modelled per polymer, so their folds carry one row per polymer with the measurements averaged, and no exposure columns.

Split granularity. The SEA and CO folds hold one row per polymer, so no test polymer appears in the training split of its fold. The tig and pkhrr folds are split by measurement row rather than by polymer: in each fold, 15 to 21 of the test polymers also appear in that fold's training split, at a different thickness or flux, or as a replicate of the same condition. A model evaluated on these folds is therefore being asked to interpolate across exposures of polymers it has seen, not to generalize to unseen polymers. Group the rows by smiles_canonicalized and re-split if you need the stricter protocol.

---

Provenance

The v1 structures are enumerated combinatorially from a library of backbone templates and pendant groups. Every candidate is validated, written with each open end as a bare asterisk, and stored as a canonical SMILES string.

The v2 labels are simulated, not measured. The pyrolysis model takes ten polymer properties — density, thermal conductivity, specific heat capacity and surface emissivity; onset of decomposition, temperature at peak mass loss rate and heat of pyrolysis; and total heat of combustion, combustion efficiency and char. Surface emissivity is held fixed at 0.88 and the remaining nine are estimated from the repeat-unit SMILES by quantitative structure–property relationships.

The experimental measurements are drawn from the published cone-calorimeter literature. The per-polymer source for each of the 49 materials is listed in Table S1 of the supplementary information accompanying the paper.

Citation

Physics-guided chemical language modelling for polymer flammability prediction and chemically meaningful inverse design.

Code: https://github.com/fishmoon1234/poly4mer · Model weights: https://huggingface.co/fishmoon1234/poly4mer

License

MIT.