Data for “Source-disjoint validation reveals transferability limits of literature-trained machine learning for Ru-based CO₂ methanation”
Description
This dataset supports the manuscript entitled “Source-disjoint validation reveals transferability limits of literature-trained machine learning for Ru-based CO₂ methanation”. The dataset contains 556 experimental records for Ru-based CO₂ methanation curated from 41 literature DOI sources. The records include catalyst composition, Ru loading, support and promoter information, BET surface area, pretreatment conditions, reaction temperature, pressure, gas hourly space velocity (GHSV), H₂/CO₂ ratio, CO₂ conversion, and CH₄ selectivity. The accompanying code implements CO₂-conversion regression and high-CH₄-selectivity classification (CH₄ selectivity ≥95%), together with conventional random-split validation, DOI-grouped cross-validation, source-disjoint external validation, and SHAP-based model interpretation. DOI identifiers are retained to enable publication-aware validation and to prevent observations from the same literature source from being distributed across training and validation folds. The repository is intended to support reproducibility of the source-aware machine-learning analysis and to facilitate further investigation of cross-publication generalization in literature-derived catalyst datasets. Associated code and screening results are provided to facilitate reproducibility and further analysis. The dataset is derived from previously published literature and is intended for research and academic use.
Files
Steps to reproduce
Download all files provided in this repository. The main dataset (ru_final.xlsx) contains the curated literature-derived data used for machine-learning model development. Run Ru_ML.py in a Python environment after setting the appropriate input and output file paths. The workflow performs data preprocessing, CO₂ conversion regression, high-CH₄-selectivity classification, model evaluation, SHAP-based interpretation, applicability-domain assessment, and virtual candidate screening. The external_validation_dataset.xlsx file is provided for additional model validation. Top_10_Candidates.xlsx and Pareto_Candidates.xlsx contain the ranked virtual-screening results. Detailed descriptions of the files, variables, model targets, and workflow are provided in README.txt.
Institutions
- Foshan UniversityGuangdong, Foshan