poly4mer v3.0.0

A chemical language model for polymer fire property prediction and inverse design, conditioned on the test conditions of the experiment (sample thickness and incident heat flux).

poly4mer maps a polymer pSMILES string to a single 768-dimensional latent vector z. From that latent it both reconstructs the pSMILES and predicts four cone-calorimeter properties, so the same representation supports forward prediction and gradient-based design in one space.

Code	https://github.com/fishmoon1234/poly4mer (branch dev)
Backbone	SMI-TED Light, frozen
Latent	768 dimensions
Max sequence	202 tokens, vocabulary 2393
Properties	time to ignition, peak heat release rate, smoke extinction area, CO yield
Conditions	sample thickness (mm), incident heat flux (kW/m²)
License	MIT

Predicted properties

Key	Property	Unit
tig	time to ignition	s
pkhrr	peak heat release rate	kW/m²
sea	smoke extinction area	m²/kg
co	carbon monoxide yield	kg/kg

Files

This repository holds both released versions. v3 lives under poly4mer_pretrained_v3/.

Path in this repository	Size	Where it goes in the code repo
poly4mer_pretrained_v3/poly4mer_v3.ckpt	3.7 GB	checkpoint/
poly4mer_pretrained_v2/smi_ted_light/smi-ted-Light_40.pt	1.1 GB	smi_ted_light/
poly4mer_pretrained_v3/tig_10folds/fold_{0..9}_best_fp32.ckpt	32 MB	poly4mer_monotone/checkpoint/tig/last_two_linear/
poly4mer_pretrained_v3/phrr_10folds/fold_{0..9}_best_fp32.ckpt	32 MB	poly4mer_monotone/checkpoint/phrr/last_two_linear/
poly4mer_pretrained_v3/sea_10folds/fold_{0..9}.ckpt	32 MB	checkpoint/fire_property_heads/sea/
poly4mer_pretrained_v3/co_10folds/fold_{0..9}.ckpt	32 MB	checkpoint/fire_property_heads/co/

Only the first two are needed for prediction, reconstruction or design. The main checkpoint already carries a property head for each of the four properties.

The per-fold folders hold two different kinds of head. For tig and pHRR they are the monotone heads described below, recognisable by the _best_fp32 suffix. For SEA and CO, which have no monotone variant, they are the standard per-fold heads used to reproduce the cross-validation tables. The v2 release is under poly4mer_pretrained_v2/.

What is in poly4mer_v3.ckpt

Key	Contents
star_encoder	embeds the * attachment-point token
autoencoder_encoder	token embeddings → latent z
decoder	latent z → token embeddings
predictor	token embeddings → token logits (the token predictor, not the property one)
property_regressors	four MLP heads for tig, pHRR, SEA, CO
property_stats	standardization statistics for thickness, flux and the four targets

Two kinds of "predictor". predictor is the token predictor, the last stage of the decoder. The property predictors live under property_regressors.

The property heads consume the raw latent. Only thickness and flux are standardized.

Head	Input	Dimension
tig	[z, thickness_norm, flux_norm]	770
pHRR	[z, thickness_norm, flux_norm]	770
SEA	z	768
CO	z	768

Each head emits a standardized value; the physical value is output * y_std[k] + y_mean[k].

Input format

Inputs must be RDKit-canonical pSMILES with bare * attachment points, for example *CC(*)c1ccccc1. Bracketed stars such as [*], and non-canonical orderings of the same molecule, do not encode faithfully and will not reconstruct.

This is the single most common cause of a failed installation reporting near-zero reconstruction. The repository ships canonical_psmiles() and a --canonicalize flag for exactly this.

The model is trained on single repeat units. Multi-component strings joined by ., used for blends and some thermoset formulations, do not reconstruct: the fragments come back but the separator does not.

Quick start

git clone -b dev https://github.com/fishmoon1234/poly4mer.git && cd poly4mer
pip install -r requirements.txt

huggingface-cli download fishmoon1234/poly4mer \
    poly4mer_pretrained_v3/poly4mer_v3.ckpt --local-dir /tmp/p4m
mv /tmp/p4m/poly4mer_pretrained_v3/poly4mer_v3.ckpt checkpoint/
huggingface-cli download fishmoon1234/poly4mer \
    poly4mer_pretrained_v2/smi_ted_light/smi-ted-Light_40.pt --local-dir /tmp/p4m
mv /tmp/p4m/poly4mer_pretrained_v2/smi_ted_light/smi-ted-Light_40.pt smi_ted_light/

python fire_property_prediction_V3.py --smiles '*CC(*)C' --thickness 3 --flux 50

>> Predicted Tig      30.0235  s
>> Predicted pHRR   2106.3787  kW/m^2
>> Predicted SEA     0.053634  m^2/kg
>> Predicted CO      0.028990  kg/kg

Inverse design from a seed polymer, and a reconstruction check:

./run_design.sh
python evaluate_reconstruction_v3.py --files homo=example_pSMILES_homo.txt

Performance

Reconstruction

Canonical pSMILES in, canonical pSMILES out, exact string match.

Set	n	Exact match	Mean Tanimoto
Training corpus sample	100,000	99.99%	1.000
Simulated set	109,993	100.00%	1.000
Commodity homopolymers	16	100.00%	1.000

The 16 homopolymers include seven absent from the training corpus; all seven reconstruct exactly, including an 87-token polyetherimide.

Property prediction

Mean absolute relative error on experimental cone-calorimeter data, 10-fold cross-validation.

Property	Best-5 folds	All 10 folds
tig	15.76%	19.11%
pHRR	12.61%	16.39%
SEA	1.58%	3.63%
CO	0.93%	3.94%

Read these with the protocol in mind. The training epoch and learning rate are selected on each fold's own test rows, so per-fold test errors are optimistically biased. Prefer the all-10 column when a single number is needed. The repository ships the fold data, the 40 per-fold heads and evaluate_fire_prediction_heads.py so both columns can be checked, plus a --cv mode that re-runs the training; re-training reproduces the best-5 figure to about 0.1 points for tig and 0.03 for CO.

Sanity against measurement, at 50 kW/m²:

Polymer	Measured pHRR	Predicted
nylon 6,6	1226	1228
polypropylene	2141	1947
PMMA	1115	1144

Intended use

Screening and ranking candidate polymer repeat units for fire behaviour, and exploring the latent neighbourhood of a known polymer. Suitable for research.

Not suitable as a substitute for cone-calorimeter testing, for safety certification, or for any decision where a wrong fire-performance estimate carries risk. The experimental data behind the property heads is small, a few hundred rows, and errors on unseen polymers are tens of percent.

Limitations

Inverse design: report the molecule, not the latent point. Properties read off the optimization path are predictions at a point in latent space, and they do not match the predictions for the molecule that point decodes to. Across a benchmark of 17 designs the median surviving fraction of the predicted gain was about one third, and four of 17 reversed sign once re-encoded. Always re-encode a candidate and predict again before quoting a number.

Design yield is modest. On 16 commodity homopolymers the loop returned about 1.1 valid designs per seed, and most were chemically implausible on inspection. Large, already fire-resistant seeds such as polyetherimides yield little, because the objective terms they already satisfy supply almost no gradient.

Weight the objective by headroom, not importance. On a seed already near optimal for a property, weighting that property heavily starves the search.

The decodable region is thin. Latents far from a real polymer decode to unparsable strings, and the property heads can return physically impossible values there, such as a negative heat release rate. A trust region on the latent displacement is advisable.

Out-of-distribution polymer databases. Reconstruction on an external polymer database is far below the figures above, around 56%, because the chemistry differs from the training distribution.

Monotone heads for tig and pHRR

tig_10folds/ and phrr_10folds/ hold alternative heads that are hard-monotone in heat flux by construction: tig can only fall as the flux rises, pHRR can only rise. The constraint is structural rather than a penalty, so it holds for every input, including polymers far outside the data. The encoder is unchanged, so these drop in beside the main checkpoint.

The head is not read out directly. The MLP is evaluated at 17 fixed flux knots from 20 to 100 kW/m², a running minimum (tig) or maximum (pHRR) is taken across them, and the value at the requested flux is interpolated between the two bracketing knots. Fluxes outside that range are clamped to the end knots.

Ten checkpoints ship per property, one per fold. --ensemble averages them, which is the safer choice for a polymer outside the experimental set.

cd poly4mer_monotone
./run_property_prediction_monotone.sh

Use them when a prediction feeds an optimizer or a screening loop and a non-physical flux trend would be costly. On held-out polymers they are level with, not better than, the standard heads under leakage-free selection, so the case for them is the guarantee rather than the accuracy. SEA and CO have no monotone variant.

Citation

@article{liu2025harnessing,
  title={Harnessing large language models for data-scarce learning of polymer properties},
  author={Liu, Ning and Jafarzadeh, Siavash and Lattimer, Brian Y and Ni, Shuna and Lua, Jim and Yu, Yue},
  journal={Nature Computational Science},
  volume={5},
  number={3},
  pages={245--254},
  year={2025},
  publisher={Nature Publishing Group US New York}
}

@inproceedings{yin2025fake,
  title={Fake It Till You Make It: Multi-Physics Synthesis Breaks the Data Barrier in Chemical Language Models},
  author={Yin, Naiyu and Liu, Ning and Chen, Jiuzhou and Lattimer, Brian Y and Lua, Jim and Yu, Yue},
  booktitle={NeurIPS 2025 Machine Learning and the Physical Sciences Workshop}
}