PolyOmics

PolyOmics is an omics-scale computational materials database containing molecular structures, simulation metadata, and diverse physical properties for more than (10^5) polymeric materials.

The database was generated primarily using RadonPy, a fully automated molecular dynamics (MD) simulation platform for polymer materials. PolyOmics was developed through a large-scale collaboration of the RadonPy Consortium to provide foundational computational data for polymer materials informatics and simulation-to-real (Sim2Real) machine learning.

The associated study is:

R. Yoshida, Y. Hayashi, et al.
Omics-scale polymer computational database transferable to real-world artificial intelligence applications
arXiv:2511.11626

Dataset DOI: 10.57967/hf/7475

---

Overview

PolyOmics provides computational data covering a broad chemical space of polymeric materials, together with molecular structures, simulation conditions, calculation metadata, and physical properties.

The database includes:

• general organic polymers,
• cellulose derivatives,
• PFAS-related polymers,
• biodegradable polyester and polycarbonate candidates,
• ladder polymers and their corresponding non-ladder structures,
• small organic molecules,
• polymer-solvent interaction parameters,
• and molecular dynamics snapshot data.

Depending on the dataset and simulation workflow, calculated properties include thermodynamic, thermal, transport, mechanical, dielectric, optical, and structural quantities.

The database is intended for applications such as:

• polymer materials informatics,
• molecular property prediction,
• pretraining of machine-learning models,
• transfer learning from simulation to experimental data,
• chemical-space exploration,
• virtual screening,
• and benchmarking of data-driven materials models.

---

Dataset organization

Main tabular datasets

File	Description
all_data.csv	Combined table containing the main PolyOmics datasets. This is the default Hugging Face configuration.
general_polymers_with_sp_abbe_dynamic-dielectric.csv	73,045 general polymers in the isotropic amorphous state, with up to 55 calculated properties including solubility parameters, Abbe number, and dynamic dielectric properties.
cellulose.csv	50,661 cellulose derivatives.
PFAS.csv	3,821 PFAS-related structures.
biodegradable_candidate_polyester.csv	357 polyesters and 26 polycarbonates synthesized as candidates for biodegradability screening.
ladder_polymers.csv	519 ladder polymers.
corresponding_non_ladder_polymers.csv	Non-ladder polymer structures corresponding to the ladder-polymer dataset.
small_molecules.csv	23,361 small organic molecules.

The individual datasets do not necessarily contain identical columns. Therefore, users interested in a particular chemical family are encouraged to use the corresponding dataset rather than assuming that every property is available for every entry in all_data.csv.

---

Chi-parameter data

The directory

chi_parameter/

contains predicted polymer-solvent interaction parameters for 93,331 polymers.

The current collection covers commonly used organic solvents and plasticizers, including files such as:

Summary_1,4-Dioxane.csv
Summary_Acetone.csv
Summary_Acetyl_tributyl_citrate.csv
Summary_Benzene.csv
Summary_Chloroform.csv
Summary_Cyclohexane.csv
Summary_DEGDB.csv
Summary_DEHA.csv
Summary_DEHP.csv
Summary_DMSO.csv
Summary_Diethyl_ether.csv
Summary_Dimethylformamide.csv
Summary_Ethanol.csv
Summary_Ethyl_acetate.csv
Summary_Hexane.csv
Summary_Methanol.csv
Summary_Tetrahydrofuran.csv
Summary_Toluene.csv
Summary_Water.csv

These values were estimated using machine-learning models and should therefore be distinguished from properties directly obtained from molecular dynamics simulations.

Because the solvent or plasticizer identity is encoded in the filename, the chi-parameter CSV files are intentionally kept as separate files rather than automatically concatenated into the default Hugging Face dataset configuration.

---

Molecular dynamics snapshots

The directory

MD_snapshot_JSON/

contains compressed archives of molecular dynamics snapshot data.

The files are grouped according to polymer classes, for example:

MD_snapshot_JSON/
├── PAMD.tar.gz
├── PANH.tar.gz
├── PARC.tar.gz
├── PCBN.tar.gz
├── PDIE.tar.gz
├── PEST.tar.gz
├── PHAL.tar.gz
├── PHYC.tar.gz
├── PI.tar.gz
├── PIMD.tar.gz
├── PIMN.tar.gz
├── PKTN.tar.gz
├── POXI.tar.gz
├── PPHS.tar.gz
├── PPNL.tar.gz
├── PSFO.tar.gz
├── PSTR.tar.gz
├── PSUL.tar.gz
├── PURA.tar.gz
├── PURT.tar.gz
└── PVNL.tar.gz

These archives are substantially larger than the tabular property datasets and are intended for users who require molecular-level simulation structures in addition to the summarized properties.

---

Data fields

The exact columns depend on the dataset and the version of the simulation workflow.

Representative categories of information include the following.

Molecular identity and structure

Examples include:

UUID
monomer_ID
smiles_list
smiles_1
smiles_2
smiles_3
smiles_4
copoly_ratio_list
copoly_type
tacticity

Where a UUID is provided, it should be used as the primary identifier rather than the row number.

---

Simulation metadata

PolyOmics records information required to track the conditions and software environment used to generate the calculations.

Representative columns include:

temp
press
input_natom
input_nchain
ini_density
forcefield

Python_ver
RadonPy_ver
RDKit_ver
Psi4_ver
LAMMPS_ver

preset_eq_ver
preset_tc_ver

The exact metadata available varies between calculations.

---

Molecular and quantum-chemical properties

Representative quantities include:

mol_weight
vdw_volume

qm_total_energy
qm_homo
qm_lumo

qm_dipole
qm_polarizability

Some quantities are stored separately for individual monomer components.

---

Equilibrium and structural properties

Examples include:

density
Rg
Scaled Rg
self-diffusion

Cp
Cv

compressibility
isentropic_compressibility

bulk_modulus
isentropic_bulk_modulus

volume_expansion
linear_expansion

nematic_order_parameter

---

Thermal properties

Examples include:

thermal_conductivity
thermal_diffusivity

Additional columns decompose the calculated thermal conductivity into contributions from different terms in the molecular mechanics Hamiltonian.

---

Glass-transition-related properties

Where the corresponding workflow was performed, the dataset may contain quantities such as:

tg
tg_rmse
tg_cooling_rate
tg_max_temp
tg_min_temp
tg_interval_temp

---

Optical and dielectric properties

Depending on the dataset, available properties include:

refractive_index
refractive_index_sos
abbe_number_sos

static_dielectric_const

efdp_permittivity_real
efdp_permittivity_imaginary
efdp_dielectric_loss_tan

---

Solubility parameters

Representative fields include:

sp_ced
sp_total
sp_vdw
sp_ele
sp_ele_short
sp_ele_long

---

Loading the dataset

Using Hugging Face Datasets

The complete tabular dataset can be loaded as:

from datasets import load_dataset

dataset = load_dataset(
    "yhayashi1986/PolyOmics",
    "all_data",
    split="train",
)

all_data is the default configuration, so the following is equivalent:

dataset = load_dataset(
    "yhayashi1986/PolyOmics",
    split="train",
)

A specific subset can be loaded by selecting its configuration:

dataset = load_dataset(
    "yhayashi1986/PolyOmics",
    "general_polymers",
    split="train",
)

Other available configurations include:

all_data
general_polymers
cellulose
pfas
biodegradable_candidate_polyesters
ladder_polymers
corresponding_non_ladder_polymers
small_molecules

Note that the Hugging Face split name train is used only as a dataset container. PolyOmics does not prescribe a particular machine-learning train/test partition. Users should construct appropriate train, validation, and test splits for their scientific application.

---

Loading an individual CSV with pandas

Individual files can be downloaded using huggingface_hub:

import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="yhayashi1986/PolyOmics",
    repo_type="dataset",
    filename="general_polymers_with_sp_abbe_dynamic-dielectric.csv",
)

df = pd.read_csv(path, low_memory=False)

print(df.shape)
print(df.columns)

---

Loading chi-parameter data

For example, the polymer-water interaction data can be accessed as:

import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="yhayashi1986/PolyOmics",
    repo_type="dataset",
    filename="chi_parameter/Summary_Water.csv",
)

df_water = pd.read_csv(path)

The same procedure can be used for the other solvents and plasticizers.

---

Downloading molecular dynamics snapshots

For example:

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="yhayashi1986/PolyOmics",
    repo_type="dataset",
    filename="MD_snapshot_JSON/PHYC.tar.gz",
)

print(path)

Because the MD snapshot archives are large, users are encouraged to download only the polymer classes required for their analysis.

---

Missing values

PolyOmics combines results from several simulation workflows that have evolved over the development of RadonPy.

Not every physical property is available for every material.

A missing value should not automatically be interpreted as a failed simulation.

Users should inspect the relevant calculation and validation fields when filtering the dataset for a particular property.

---

Reproducibility and versioning

PolyOmics is a continuously evolving computational database.

Simulation metadata, including RadonPy, RDKit, Psi4, LAMMPS, and workflow versions, are included in the dataset where available to facilitate reproducibility.

For analyses intended for publication, we recommend recording the exact Hugging Face repository revision used.

For example:

dataset = load_dataset(
    "yhayashi1986/PolyOmics",
    "general_polymers",
    split="train",
    revision="<commit-hash>",
)

This is particularly important because additional structures, calculated properties, or recalculated records may be released in future versions.

---

Important considerations and limitations

PolyOmics is primarily a computational database, not an experimental property database.

The values should therefore be interpreted within the assumptions of the corresponding molecular simulation and quantum-chemical workflows.

Important considerations include:

1. Force-field dependence
   Molecular dynamics results depend on the molecular mechanics force field and parameterization procedure.

2. Finite simulation time and system size
   Calculated values can depend on sampling time, simulation-cell size, and equilibration quality.

3. Idealized morphology
   Many polymer calculations represent isotropic amorphous systems. Real polymer materials can additionally contain crystallinity, orientation, processing history, interfaces, additives, defects, and other mesoscale structures.

4. Stereoregularity
   The representation of tacticity and stereoregularity may be simplified for some entries and should be considered when comparing PolyOmics values with experimental materials.

5. Simulation-to-experiment discrepancy
   Calculated properties should not be interpreted as certified experimental material performance. One intended use of PolyOmics is precisely to learn transferable representations from large-scale simulations and adapt them to experimental data using approaches such as transfer learning.

6. Heterogeneous property coverage
   Different subsets and workflow generations contain different sets of calculated properties.

Users conducting quantitative comparisons with experimental measurements should consider these sources of systematic and statistical uncertainty.

---

Recommended use for machine learning

PolyOmics is particularly suitable as a large-scale source-domain dataset for polymer machine learning.

A typical workflow is:

PolyOmics computational data
        │
        ▼
pretraining / representation learning
        │
        ▼
small experimental target dataset
        │
        ▼
fine-tuning / transfer learning
        │
        ▼
prediction of real-world polymer properties

The associated PolyOmics study demonstrates that the generalization performance of simulation-to-real transfer models can improve systematically as the amount of computational pretraining data increases.

---

Citation

If you use PolyOmics in academic work, please cite both the dataset and the associated PolyOmics paper.

Dataset

PolyOmics
RadonPy Consortium
Hugging Face Dataset
DOI: 10.57967/hf/7475

Paper

Ryo Yoshida, Yoshihiro Hayashi, et al.
Omics-scale polymer computational database transferable to real-world artificial intelligence applications
arXiv:2511.11626 (2025)

A minimal BibTeX entry for the paper is:

@article{yoshida2025polyomics,
  title   = {Omics-scale polymer computational database transferable
             to real-world artificial intelligence applications},
  author  = {Yoshida, Ryo and Hayashi, Yoshihiro and others},
  journal = {arXiv preprint arXiv:2511.11626},
  year    = {2025},
  doi     = {10.48550/arXiv.2511.11626}
}

---

License

© 2026 The RadonPy Consortium.

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Redistribution and reuse are permitted under the terms of CC BY 4.0. Please provide appropriate attribution when using or redistributing the dataset.

---

Contact and contributions

PolyOmics is developed and maintained by the RadonPy Consortium.

Questions, bug reports, data-related issues, and suggestions can be submitted through the Community section of this Hugging Face dataset repository.

We also welcome scientific use of PolyOmics for benchmarking, materials informatics, simulation-to-experiment transfer learning, and the development of new polymer informatics methods.