Unlocking prediction and optimal design of CO2 methanation catalysts via active learning-enhanced interpretable ensemble learning

Proposed a framework for predicting and optimizing CO2 methanation catalysts using active learning-enhanced interpretable ensemble learning. [Display omitted] •An active learning-enhanced ensemble framework is proposed for catalyst optimization.•Shapley additive explanations and partial dependence a...

Celý popis

Uložené v:
Podrobná bibliografia
Vydané v:Chemical engineering journal (Lausanne, Switzerland : 1996) Ročník 509; s. 161154
Hlavní autori: Yang, Qingchun, Bao, Runjie, Wang, Zhao, Guo, Qiwen, Pan, Yifei, Zhou, Xin
Médium: Journal Article
Jazyk:English
Vydavateľské údaje: Elsevier B.V 01.04.2025
Predmet:
Active learning
Catalyst optimization
CO2 methanation
Ensemble models
Interpretable
Ensemble models
CO2 methanation
Active learning
Interpretable
Catalyst optimization
ISSN:1385-8947
On-line prístup:Získať plný text
Tagy: Pridať tag
Žiadne tagy, Buďte prvý, kto otaguje tento záznam!
Popis
Shrnutí:Proposed a framework for predicting and optimizing CO2 methanation catalysts using active learning-enhanced interpretable ensemble learning. [Display omitted] •An active learning-enhanced ensemble framework is proposed for catalyst optimization.•Shapley additive explanations and partial dependence analysis is conducted to interpret the model.•Uncertainty sampling strategy is the most effective in enhancing the performance of ensemble learning models.•Active learning-optimized random forest model is the most suitable for the CO2 methanation reaction.•Optimized Ni/Al2O3 and Ni/CeO2 catalysts achieve CH4 yields of 92.12 % and 92.72 %. CO2 methanation technology effectively promotes the recycling of carbon resources and sustainable development. However, its intricate reaction process presents significant challenges for catalyst design and optimization. Herein, an efficient, robust, and interpretable ensemble learning model based on the active learning optimization strategy is developed to optimize and reverse design highly active catalysts for CO2 methanation. First, various feature engineering methods were compared and it found feature selection based on recursive feature elimination and cross-validation is the most suitable approach for the CO2 methanation process, outperforming the optimal feature extraction method, Autoencoders. Six ensemble learning models are then developed and automatically optimized using the Optuna framework. To enhance the prediction accuracy and generalization ability of the model, various active learning strategies are devised, with findings indicating that the uncertainty sampling strategy significantly improves the performance of ensemble learning models. In particular, the active learning optimized random forest model exhibits superior performance due to its highest R2 value (>0.92) and the most reasonable prediction range. To clearly explain the outstanding performance of the model, the Shapley additive explanations and partial dependence plot analyses are conducted to effectively illustrate the significance of each feature in predicting outcomes and elucidating their interrelationships. The optimal model is ultimately integrated with a hybrid multi-objective algorithm to optimize the reported catalysts, and successfully identified Ni/Al2O3 and Ni/CeO2 catalysts with methane yields of 92.12 % and 92.72 %, surpassing previously reported data. Additionally, the model also predicted three novel catalysts with superior performance at low temperatures and varying H2/CO2 ratios.
ISSN:1385-8947
DOI:10.1016/j.cej.2025.161154