A novel machine learning framework for designing high-performance catalysts for production of clean liquid fuels through Fischer-Tropsch synthesis

• Published in: Energy (Oxford) Vol. 289; p. 130061
• Main Authors: Teimouri, Zahra, Abatzoglou, Nicolas, Dalai, Ajay K.
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 15-02-2024
• Subjects: Carbon supports; Catalyst design; Clean transportation fuels; Fischer-Tropsch synthesis; Machine learning; Fischer-Tropsch synthesis; Catalyst design; Carbon supports; Clean transportation fuels; Machine learning
• ISSN: 0360-5442
• DOI: 10.1016/j.energy.2023.130061

Fischer-Tropsch synthesis (FTS) attracts great interest as a sustainable route for production of sustainable transportation fuels. Catalyst design and operational conditions tune the selectivity of FTS to liquid fuels, which can be predicted with machine learning models (ML). Herein, a comprehensive dataset including 27 key input features related to the catalyst structure (with the focus on carbon supports), preparation, activation, and FTS operating conditions were investigated to predict the CO conversion and C5+ selectivity using three ML algorithms. Feature engineering and selection were implemented to identify the significant catalyst formulation descriptors. Principal component analysis was used to explore the information space and decrease the dimension of the dataset before ML prediction. In addition to the well-known effects of the operational conditions on FTS, roles of physico-chemical properties of the carbon materials were also extensively analyzed. Random forest (RF) including 4 principal components, indicated the highest accuracy for prediction of the CO conversion and C5+ selectivity with R2 of 0.91 and 0.97, respectively. The proposed framework can provide a reliable strategy in designing efficient carbon supported catalysts for FTS and guide experiments by identifying the key descriptors in the catalyst formulation and operating conditions to enhance the selectivity of liquid fuels.
[Display omitted]
•A machine learning framework for CO hydrogenation to clean liquid fuels with the focus on carbon supports was developed.•Catalyst characteristics, preparation, activation, and operating conditions were compiled in the dataset.•Principal component analysis was conducted to evaluate the information space before machine learning modeling.•Random forest led to the highest prediction accuracy of catalyst selectivity to C5+ with R2 = 0.97.•According to the SHAP theory, catalyst characteristics were identified as the most significant features.