PolyCARP predicts copolymer architecture from two monomers and a reaction recipe. It combines machine learning with literature lookup to distinguish alternating, random, and gradient regimes. The underlying dataset contains 3,791 reactions from 1,206 publications.
Use the web tool · Model card · Installation and demo
With Git and uv installed:
git clone https://github.com/lamalab-org/PolyCARP.git
cd PolyCARP
uv sync --locked --python 3.12.11 --extra reproduction
uv run --locked --extra reproduction python 02-reactivity-prediction/copol_prediction/reproduce_paper_metrics.pyThe demo uses bundled data, descriptors, and weights. It checks the paper's train/test table to a tolerance of 0.005 and exits nonzero on a mismatch. See the reproduction guide for tested versions, hardware, measured install and demo times, expected output, and use with your own data.
The extraction pipeline turns literature into measurements. The prediction pipeline uses those measurements to train the model and serve recipe predictions.
| Directory | Contents |
|---|---|
| 01-data-extraction | Literature search, PDF extraction, and curation |
| 02-reactivity-prediction/copol_prediction | Descriptors, training, evaluation, and analysis |
| api | Local API and Docker setup; web interface source |
| artifacts | Released model, train/validation/test splits, and metrics |
| paper_dataset | Frozen dataset used in the paper |
| 03-experiments | Baselines, ablations, and laboratory and solvent case studies |
| 04-NOMAD-database | Dataset-summary figures |
| src | copolpredictor and copolextractor Python libraries |
Install optional dependencies with uv sync --locked --extra NAME, where NAME
is extraction, training, testing, or database. Combine extras as needed.
Run scripts with the same extras, for example:
uv run --locked --extra testing pytest tests@article{SchillingWilhelmi2026,
title = {Condition-aware prediction of copolymer architecture},
url = {http://dx.doi.org/10.26434/chemrxiv.15004102/v2},
DOI = {10.26434/chemrxiv.15004102/v2},
publisher = {ChemRxiv},
author = {Schilling-Wilhelmi, Mara and Bulgakov, Boris and Patiny, Luc and Kapoor, Sarthak and Jablonka, Kevin Maik},
year = {2026},
month = June
}MIT — see LICENSE.