Data-Driven Exploration of Literature-Derived Catalyst and Reaction Spaces for the Vapor-Phase Aldol Condensation of Acetate with Formaldehyde
- 1Centre for Sustainable and Catalysis Engineering (CSCE), KU Leuven, Celestijnenlaan 200F, Leuven, Belgium, 3001
Abstract
Emerging catalytic processes are often characterized by sparse, heterogeneous, literature-derived datasets, limiting the applicability of conventional predictive machine learning. Herein, we demonstrate a robust approach for extracting mechanistic insights from such challenging datasets using interpretable machine learning. A database comprising 375 observations from 71 publications on the vapor-phase thermocatalytic aldol condensation of acetate with formaldehyde to acrylate was compiled, including catalyst composition, reaction conditions, and catalytic performance. Unsupervised clustering identifies three catalyst families: titanium and vanadium phosphorus oxides, alumina-supported Cs- and Ba-based catalysts, and silica-supported Cs-based catalysts. The small dataset size, restricted catalyst diversity, and inter-study variability result in substantial model overfitting and constrain the accurate prediction of catalyst performance. Robust model development relies on regularized tree-based estimators, publication-grouped cross-validation, and sensitivity analysis across model architectures and random initializations. SHAP analysis identifies liquid hourly space velocity and reactant molar ratio as the dominant operational variables governing acrylate productivity. The presence of water, methanol, or ester-based acetate precursors lowers productivity, while reaction temperature exhibits limited influence. Metal phosphorus oxide and silica-supported Cs materials are ranked as the most and least active catalyst families, respectively. Despite catalyst deactivation being the major limitation, time-on-stream profiles remain largely underreported. Correlation analysis associates the deactivation of metal phosphorus oxide and alumina-supported catalysts primarily with acrylate productivity, consistent with acrylate polymerization, whereas silica-supported Cs catalysts exhibit greater variability, suggesting additional deactivation pathways. Overall, this work provides a transferable framework for extracting reliable insights from sparse literature-derived datasets using interpretable machine learning and robust validation.
Related
- Construction of a Curated Literature Dataset and Machine Learning-Guided Exploration of Supported Metal Catalysts for Low- to Intermediate-Temperature Reverse Water–Gas Shift Reaction
- Bridging the Literature-to-Laboratory Gap with Language Models and Active Learning for Reaction Feasibility Prediction and Autonomous Exploration
- Microwave electric field-induced dynamic hydration enables selected dehydration of formic acid over a poly(m-phenolsulfonic acid–formaldehyde Catalyst
- Atmospheric pressure water-gas shift reaction and transfer hydrogenation of ketones using a ruthenium pincer catalyst.
- Challenges in Catalyst Design for the Ammonia Oxidation Reaction in Aqueous Media: The Case of Metal Porphyrins
BibTeX
@misc{verstraeten2026datadriven,
title = {{Data-Driven Exploration of Literature-Derived Catalyst and Reaction Spaces for the Vapor-Phase Aldol Condensation of Acetate with Formaldehyde}},
author = {Simon B. Verstraeten and Yayati Naresh Palai and Ekaterina V. Makshina and Bert F. Sels},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008977/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008977/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Data-Driven Exploration of Literature-Derived Catalyst and Reaction Spaces for the Vapor-Phase Aldol Condensation of Acetate with Formaldehyde
AU - Simon B. Verstraeten
AU - Yayati Naresh Palai
AU - Ekaterina V. Makshina
AU - Bert F. Sels
PY - 2026
DA - 2026/09/17
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008977/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008977/v1
AB - Emerging catalytic processes are often characterized by sparse, heterogeneous, literature-derived datasets, limiting the applicability of conventional predictive machine learning. Herein, we demonstrate a robust approach for extracting mechanistic insights from such challenging datasets using interpretable machine learning. A database comprising 375 observations from 71 publications on the vapor-phase thermocatalytic aldol condensation of acetate with formaldehyde to acrylate was compiled, including catalyst composition, reaction conditions, and catalytic performance. Unsupervised clustering identifies three catalyst families: titanium and vanadium phosphorus oxides, alumina-supported Cs- and Ba-based catalysts, and silica-supported Cs-based catalysts. The small dataset size, restricted catalyst diversity, and inter-study variability result in substantial model overfitting and constrain the accurate prediction of catalyst performance. Robust model development relies on regularized tree-based estimators, publication-grouped cross-validation, and sensitivity analysis across model architectures and random initializations. SHAP analysis identifies liquid hourly space velocity and reactant molar ratio as the dominant operational variables governing acrylate productivity. The presence of water, methanol, or ester-based acetate precursors lowers productivity, while reaction temperature exhibits limited influence. Metal phosphorus oxide and silica-supported Cs materials are ranked as the most and least active catalyst families, respectively. Despite catalyst deactivation being the major limitation, time-on-stream profiles remain largely underreported. Correlation analysis associates the deactivation of metal phosphorus oxide and alumina-supported catalysts primarily with acrylate productivity, consistent with acrylate polymerization, whereas silica-supported Cs catalysts exhibit greater variability, suggesting additional deactivation pathways. Overall, this work provides a transferable framework for extracting reliable insights from sparse literature-derived datasets using interpretable machine learning and robust validation.
ER -