An interpretable CatBoost–Grey Wolf Optimizer framework for cross-study modeling of COF/MOF photocatalytic pesticide degradation

Covalent organic frameworks and metal-organic frameworks have been proposed as effective photocatalysts for the degradation of pesticides, yet large variations in catalyst chemistry, reaction conditions, and reporting within and across individual research articles have thus far precluded the development of universal, quantitative design rules. Herein, a harmonized dataset is collected and curated for 644 records published across eight peer-reviewed studies for a selection of 29 experimental condition-catalyst material combinations (18 parent framework designs, nine pesticides and one binary pesticide mixture) and supplemented with structure-derived molecular descriptors for the pesticides. A gradient-boosting CatBoost regression model is trained, guided by the Grey Wolf Optimizer tuning algorithm, for the prediction of normalized residual concentration of the target pollutant (C/C 0 ) as a function of operating parameters and catalyst-level features. The refined model gives a pooled out-of-fold R 2 of 0.939 (RMSE = 0.073, MAE = 0.045) under repeated random five-fold cross-validation, which reflects interpolation along the degradation curves of the compiled experiments. However, when entire experiments are held out, R 2 drops to 0.670 (RMSE = 0.171). Leave-one-catalyst-out and leave-one-study-out validation did not outperform a training-mean baseline overall (pooled R² = −1.14 and −0.53), so prediction for unseen catalysts is not supported. Shapley additive explanation analysis reveals that treatment time, oxidant/additive usage conditions, and solution pH are paramount explanatory factors, whereas pollutant molecular descriptors play a comparatively minor role. An auxiliary catalyst-ranking model that includes catalyst identity reproduces four within-study catalyst contrasts that are part of its training data, but these contrasts are not recovered when the studies are held out. Symbolic regression yields a single-variable hyperbolic expression of irradiation time that is equivalent to a conventional second-order-type kinetic form and adds no predictive information beyond it. Overall, the model provides an interpretable analysis of the compiled dataset and may help formulate hypotheses for further experiments, whereas prediction or optimization beyond the studied experimental systems requires additional data and independent validation.

Authors

Institutions

Publication Details

Journal
Next Materials
Published
2026-11-16
DOI
https://doi.org/10.1016/j.nxmate.2026.104070
Primary Topic
Advanced Photocatalysis Techniques
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

An interpretable CatBoost–Grey Wolf Optimizer framework for cross-study modeling of COF/MOF photocatalytic pesticide degradation

Mansour Ghaffari‐Moghaddam, Mostafa Khajeh
3 links
Next Materials
Advanced Photocatalysis Techniques
article

An interpretable CatBoost–Grey Wolf Optimizer framework for cross-study modeling of COF/MOF photocatalytic pesticide degradation

Mansour Ghaffari‐Moghaddam, Mostafa Khajeh
article en

Abstract

Covalent organic frameworks and metal-organic frameworks have been proposed as effective photocatalysts for the degradation of pesticides, yet large variations in catalyst chemistry, reaction conditions, and reporting within and across individual research articles have thus far precluded the development of universal, quantitative design rules. Herein, a harmonized dataset is collected and curated for 644 records published across eight peer-reviewed studies for a selection of 29 experimental condition-catalyst material combinations (18 parent framework designs, nine pesticides and one binary pesticide mixture) and supplemented with structure-derived molecular descriptors for the pesticides. A gradient-boosting CatBoost regression model is trained, guided by the Grey Wolf Optimizer tuning algorithm, for the prediction of normalized residual concentration of the target pollutant (C/C 0 ) as a function of operating parameters and catalyst-level features. The refined model gives a pooled out-of-fold R 2 of 0.939 (RMSE = 0.073, MAE = 0.045) under repeated random five-fold cross-validation, which reflects interpolation along the degradation curves of the compiled experiments. However, when entire experiments are held out, R 2 drops to 0.670 (RMSE = 0.171). Leave-one-catalyst-out and leave-one-study-out validation did not outperform a training-mean baseline overall (pooled R² = −1.14 and −0.53), so prediction for unseen catalysts is not supported. Shapley additive explanation analysis reveals that treatment time, oxidant/additive usage conditions, and solution pH are paramount explanatory factors, whereas pollutant molecular descriptors play a comparatively minor role. An auxiliary catalyst-ranking model that includes catalyst identity reproduces four within-study catalyst contrasts that are part of its training data, but these contrasts are not recovered when the studies are held out. Symbolic regression yields a single-variable hyperbolic expression of irradiation time that is equivalent to a conventional second-order-type kinetic form and adds no predictive information beyond it. Overall, the model provides an interpretable analysis of the compiled dataset and may help formulate hypotheses for further experiments, whereas prediction or optimization beyond the studied experimental systems requires additional data and independent validation.

Next MaterialsVol. 14
Zabol University (IR)
University of Zabol
Advanced Photocatalysis Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon