LubCast is an extensible platform for lubricant-property prediction. It connects two research workflows:
- experimental researchers can use existing models, train models from their own tables, compare candidates, and read reliability and interpretation reports;
- computational and machine-learning contributors can register datasets, molecular representations, estimators, complete model packages, and benchmark protocols through stable interfaces.
The platform currently supports molecular descriptors, fingerprints and molecular graphs; fold-fitted feature processing; grouped evaluation; prediction intervals; applicability-domain summaries; explanation-stability analysis; model packaging; command-line, Python and web entry points; and metadata-backed model delivery.
| Path | Content |
|---|---|
src/lubcast/ |
Core data, representation, modeling, reliability, registry and service code |
web/ |
Researcher-facing web application assets |
tests/ |
Unit, contract and workflow tests using synthetic or public fixtures |
tools/ |
Data preparation, benchmark, verification and deployment utilities |
configs/ |
Public base-oil task and benchmark examples |
data/base_oil/ |
Experimental properties for 75 ester base oils |
data/vii/ |
Published viscosity-index-improver simulation data and monomer SMILES mapping |
results/workflows/ |
Figure source tables, predictions, feature traces and plotting code for public tasks |
results/summary/ |
Compact task and benchmark summaries |
conda env create -f environment.yml
conda activate LubCast
pip install -e .
lubcast --versionOptional graph, boosting and symbolic-regression dependencies are documented in the corresponding lock files.
lubcast list-representations
lubcast list-estimators
python -m pytest -qUser workflows and contributor interfaces are documented under docs/users/ and docs/contributors/.
data/base_oil/ester-exp-75-clean.xlsxcontains the 75-molecule ester-base-oil data associated with the published study, DOI: 10.1016/j.triboint.2023.108381.data/vii/VIIInfo.csvanddata/vii/VII_SMILES.csvcontain the viscosity-index-improver data associated with the published study, DOI: 10.1038/s41524-025-01539-z.results/workflows/contains the process-level source data required to inspect the public base-oil and VII workflow figures.
The 13,320-record Fe3C adsorption-energy dataset is unpublished and is not included. Record-level splits, predictions, descriptor values, SHAP values, cached representations and fitted models derived from that dataset are also excluded. Aggregate benchmark rows retained in results/summary/ do not contain molecular structures or record-level targets.
The software interfaces remain able to accept a future adsorption-energy dataset through the same dataset and task specifications after that resource is released.
LubCast is research software. Predictions should be interpreted with the reported grouped validation, prediction interval, structural similarity and explanation-stability information. A saved estimator alone is insufficient to reproduce a screening decision; complete model packages retain preprocessing, task definition, units, data version and evaluation metadata.