DOI: 10.3303/CET24114098 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Paper Received: 23 May 2024; Revised: 28 September 2024; Accepted: 8 November 2024 
Please cite this article as: Shafiee P., Dorneanu B., Arellano-Garcia H., 2024, Towards Machine Learning-Driven Catalyst Design and 
Optimization of Operating Conditions for the Production of Jet Fuel Via Fischer-Tropsch Synthesis, Chemical Engineering Transactions, 114, 
583-588  DOI:10.3303/CET24114098 
  

 CHEMICAL ENGINEERING TRANSACTIONS  
 

VOL. 114, 2024 

A publication of 

 

The Italian Association 
of Chemical Engineering 
Online at www.cetjournal.it 

Guest Editors: Petar S. Varbanov, Min Zeng, Yee Van Fan, Xuechao Wang 

Copyright © 2024, AIDIC Servizi S.r.l. 

ISBN 979-12-81206-12-0; ISSN 2283-9216 

Towards Machine Learning-driven Catalyst Design and 

Optimization of Operating Conditions for the Production of Jet 

Fuel Via Fischer-Tropsch Synthesis 

Parisa Shafiee, Bogdan Dorneanu, Harvey Arellano-Garcia* 

Department of Process and Plant Technology. Brandenburg University of Technology (BTU) Cottbus-Senftenberg, Germany 

arellano@b-tu.de 

Fischer-Tropsch synthesis (FTS) offers a promising route for producing sustainable jet fuels from syngas. 

However, optimizing the catalyst design and operating conditions to maximize the desired C8-C16 jet fuel range 

is a challenging task. This study introduces the application of a machine learning (ML) framework to guide the 

design of Co/Fe-supported FTS catalysts and operating conditions for enhanced fuel selectivity. A 

comprehensive dataset was constructed with 21 input features spanning catalyst structure, preparation method, 

activation procedure, and FTS operating parameters. The random forest ML algorithm was evaluated for 

predicting CO conversion and C8-C16 selectivity using this dataset. Feature engineering identified the most 

significant descriptors influencing performance. A principal component analysis reduced the dataset 

dimensionality prior to ML modelling. The random forest algorithm achieved high prediction accuracy for the 

conversion of CO (R2 = 0.92) and C8-C16 selectivity (R2 = 0.90). In addition to confirming the known effects of 

operating conditions, key roles of Co/Fe-supported properties were elucidated. This ML framework provides a 

powerful tool for the rational design of FTS catalysts and operating windows to maximize jet fuel productivity. 

1. Introduction 

The growing global energy demand and concerns over climate change have spurred intense research into 

developing renewable and sustainable energy resources. Fischer-Tropsch synthesis (FTS), a catalytic process 

that converts syngas – a mixture of hydrogen and carbon monoxide (CO + H2) into light olefins, liquid fuels, and 

value-added chemicals, has received renewed attention in recent decades (Grainca et al., 2023). Syngas can 

be derived from sustainable processes such as biomass gasification, making FTS an attractive route for 

producing clean transportation fuels and chemicals (Sineva et al., 2023). The product distribution in the FTS 

process is heavily influenced by the catalyst composition and formulation, as well as the reaction conditions 

employed. Cobalt and iron are two of the most widely utilized active metal components for catalyzing the 

hydrogenation of syngas in FTS catalysts (Kriventceva et al., 2013). 

The design of an FTS catalyst involves several critical factors, such as the catalyst preparation method, the 

choice of support material, the addition of promoters, and the calcination and reduction treatments applied. Due 

to the complex nature of FTS, which involves multiple interconnected reaction steps, accurately predicting the 

product selectivity poses a significant challenge. In this context, modeling techniques offer an alternative 

approach to tackle these difficulties by enabling the prediction of optimal catalyst structures and operating 

conditions tailored to the desired product distribution (Alayat et al., 2019). Machine learning (ML) methodologies 

have found widespread application in deciphering the intricacies of many industrial applications (Dorneanu et 

al., 2022), including catalyst screening (Tomacruz et al., 2022) or catalytic reactions characterized by complex 

reaction mechanisms, such as the FTS process. However, the majority of prior studies in this domain have 

concentrated either on specific catalyst formulations or on operating conditions, failing to adopt a holistic 

approach that simultaneously considers the effects of both catalyst structure and process parameters. Notably, 

the influence of catalyst support materials, particularly zeolites, Al2O3, and SiO2, on the FTS reaction has not 

been extensively investigated using ML techniques.  

583



One of the challenges in investigating the influence of catalyst support materials on the FTS process towards 

jet fuel production using ML techniques is limited availability of high-quality experimental data. ML models 

require large datasets for training, but there is a scarcity of comprehensive experimental data on the 

performance of various support materials for jet fuel production via the FTS process. Most studies focus on 

specific support materials or operating conditions, making it difficult to develop generalizable models. However, 

with the increasing availability of experimental data, advancements in computational resources, and 

interdisciplinary collaborations, machine learning is expected to play a more significant role in optimizing catalyst 

support materials for the FTS process towards jet fuel production in the future. This represents a significant gap 

in the current understanding of the process, as these support materials can potentially play a crucial role in 

modulating the reaction mechanism and product selectivity (Teimouri et al., 2021). Zeolites, Al2O3, and SiO2 

possess unique structural and textural properties that can significantly influence the FTS reaction mechanism 

and product selectivity towards jet fuels. These support materials can modulate metal-support interactions, alter 

the electronic properties of the active phase, and introduce acidity or basicity, thereby affecting the product 

distribution. The choice of support material and its characteristics, such as specific surface area, pore size, and 

surface acidity, play a crucial role in tuning the selectivity toward jet fuels in FTS.  

Previous studies have explored using ML approaches to investigate product distribution in the FTS process. 

However, a comprehensive model that simultaneously considers the effects of operational conditions and 

catalyst structure, including catalyst synthesis, reduction, and the FTS process itself, specifically focusing on 

zeolites, Al2O3, and SiO2-based supports towards jet fuels, has been lacking (Sun et al., 2011). Teimouri et al. 

(2024) developed an ML framework to predict the performance of carbon-supported Fe/Co catalysts for FTS of 

clean liquid fuels. Their research offers a strategic plan for engineering highly effective catalysts derived from 

carbon materials. Essentially, their study provides guidelines for the systematic development of carbon-based 

catalytic systems geared towards producing renewable or non-petroleum-based fuels for various modes of 

transportation. Thus, this study aims to develop a comprehensive ML framework to predict the FTS performance 

of zeolite, alumina, and silica-supported catalysts towards jet fuel production by considering the simultaneous 

effects of catalyst structure, preparation method, activation procedure, and operating conditions. Ultimately, this 

work will provide valuable insights for the rational design of efficient zeolite, alumina, and silica-based catalysts 

tailored for specific FTS product streams towards jet fuel. 

2. Methodology 

Data curation is a critical step in developing an effective ML algorithm, as the quality of the input data significantly 

affects the model's performance. This work curated a dataset from 10 recent publications (comprising 43 

datasets) that investigated the use of zeolite, alumina, and silica supports for Fischer-Tropsch synthesis (FTS) 

catalysts toward jet fuel production. The compiled dataset was first statistically analyzed to ensure its robustness 

and then mechanistically evaluated to understand the underlying relationships between the input features and 

the desired output. Furthermore, 21 relevant descriptors related to catalyst formulation and operational 

conditions were identified as input features for the ML models. Given the importance of producing sustainable 

jet fuels via FTS, the C8-C16 selectivity (jet fuel range) and CO conversion were set as the target responses. 

Subsequently, these 21 input features were classified into two main categories: catalyst formulation and 

operational conditions. The catalyst formulation descriptors included the support material (zeolite, alumina, or 

silica), active metal (Co or Fe), promoter type and loading, catalyst preparation method, calcination temperature, 

and textural properties (surface area and pore size). The operational condition descriptors encompassed 

parameters related to catalyst reduction (temperature, pressure, and gas composition), and FTS reaction 

conditions (temperature, pressure, H2/CO ratio, and space velocity).  

 

 

Figure 1: The stages of the ML framework 

To ensure the quality and completeness of the dataset, rigorous data cleaning and preprocessing steps were 

performed, as illustrated in Figure 1. These steps included handling missing values through appropriate 

584



imputation techniques, identifying and addressing outliers, and applying feature scaling and encoding to ensure 

consistent data representation for the ML algorithms. 

3. Results 

3.1 Analysis of input data 

The relationships between input features (catalyst descriptors and operational conditions) and output variables 

(C8-C16 selectivity, CO conversion) were analyzed using Spearman rank correlation and principal component 

analysis (PCA) (Figure 2). Positive correlations were found between reduction atmosphere and active metal 

type (Fe requires H2/CO, Co requires H2), as well as between larger catalyst pore size and improved C8-C16 

selectivity due to better diffusion. A negative correlation was seen between promoter loading and C8-C16 

selectivity, suggesting that over-promotion can block active sites. PCA revealed the first 6 principal components 

accounted for over 80 % of variance, with active metal/promoter in PC1, reduction conditions in PC2, catalyst 

synthesis conditions in PC4, and FTS reaction conditions in PC5, highlighting the importance of both catalyst 

formulation and operational parameters (Figure 3). An examination of the experimental data projected onto the 

first two principal components revealed some key insights (Figure 4). Most data points clustered along a linear 

plane, reflecting variations from changing operational conditions. However, some distinct outlier clusters were 

identified, corresponding to unique reaction conditions differing significantly from the rest of the dataset. 

Visualizing the principal component projections allowed for the identification of patterns, groupings, and 

anomalies within the data. A scatter plot of PCA1 vs PCA2, color-coded by active metal type, showed distinct 

clusters for Fe vs. Co catalysts, indicating that Fe has a significant influence on the variance captured by PCA1. 

The projections also revealed unexplored regions in the parameter space, highlighting gaps in the existing 

literature and need for further research to optimize jet fuel production by investigating these unexplored areas. 

 

Figure 2: Spearman correlation matrix showing the relationships between input features and output responses 

585



 

Figure 3: Principal component analysis of the dataset: correlation of the input features and responses with all 

principal components 

 

Figure 4: PCs categorized by type of active metal 

3.2 Optimization of the machine learning model 

In the next step, the performance of a random forest ML model in accurately predicting the CO conversion rate 

and selectivity for C8-C16 hydrocarbons of a Fischer-Tropsch synthesis (FTS) catalyst under varying formulations 

and operating parameters (Figure 5) has been evaluated. The random forest model underwent optimization by 

adjusting the number of decision trees and their maximum depth. The optimal configurations were identified as 

10 trees with a maximum depth of 6 for CO conversion, and 6 trees with a maximum depth of 6 for C8-C16 

selectivity. The refined model displayed remarkable prediction accuracy, yielding average R2 values of 0.92 

(training) and 0.72 (testing) for CO conversion, and 0.90 (training) and 0.70 (testing) for C8-C16 selectivity. 

Despite a slight decline in prediction accuracy on the test dataset compared to the training data, the overall 

performance was deemed sufficient for effectively extrapolating the FTS activity and selectivity across different 

catalysts and operating conditions. Furthermore, the study evaluated the model's performance with and without 

utilizing PCA for dimensionality reduction. The findings indicate that the random forest model without PCA may 

perform better in directly capturing the FTS performance. 

586



(a) 

 

(b) 

 

(c) 

 
(d) 

 

(e) 

 

(f) 

 

Figure 5: (a) RF hyperparameter tuning for CO conversion, (b) and (c) comparison of predicted CO conversion 

with original experimental data, (d) RF hyperparameter tuning for C8-C16 selectivity, and (e) and (f) comparison 

of predicted C8-C16 selectivity with original experimental data 

3.3 Impacts of the input features on catalyst activity and selectivity 

The pie chart illustrated in Figure 6 reveals that catalyst preparation conditions (type of active metal, promoter, 

promoter loading) were identified as the least significant features influencing the catalytic activity compared to 

catalyst characterization and FTS operational conditions. 

  

(a)  

 

(b) 

 

Figure 6: Effects of input features on (a) CO conversion, (b) C8-C16 selectivity 

(a) 

 

(b) 

 

Figure 7: Two-way PDPs of (a) CO conversion and (b) C8-C16 selectivity 

587



While FTS operational conditions predominantly affect the CO conversion, catalyst characterization plays a 

more significant role in determining the C8-C16 selectivity of the catalysts. The two-way partial dependence plots 

(PDPs) illustrated in Figure 7 further validate the high interactions and non-linear relationship between the 

temperature and pressure of the reaction, and catalyst activity. Notably, within the pressure range of 12.5-20 

bar, higher CO conversion was observed at temperatures that exceed 250 °C, as depicted in Figure 2. These 

plots collectively underscore the dominance of catalyst characterization parameters; particularly support acidity 

and specific surface area, in determining the selectivity towards C8-C16 hydrocarbons. These findings highlight 

the necessity of tailoring catalyst properties to achieve the desired product distribution in the FTS process. 

Specifically, the utilization of zeolite supports with engineered textural properties and surface chemistry is crucial 

for designing catalysts with moderate acidity and optimum surface area. By carefully controlling these support 

characteristics, it becomes feasible to finely adjust the catalyst's selectivity towards the desired C8-C16 

hydrocarbon range. 

4. Conclusions 

This study developed a machine learning (ML) framework utilizing a random forest (RF) model to predict the 

performance of Fe/Co catalysts in producing jet fuels (C8-C16) through the Fischer-Tropsch synthesis (FTS) 

process. The RF model demonstrated high prediction accuracy for CO conversion (R2 of 0.92 train, 0.72 test) 

and C8-C16 selectivity (R2 of 0.90 train, 0.70 test). Partial dependence plots revealed coupling effects between 

catalyst textural properties, synthesis conditions, and FTS reaction conditions on the efficiency of the process. 

This framework offers insights for designing improved FTS catalysts for selective jet fuel production and holds 

potential for expansion by incorporating additional catalyst descriptors. Moving forward, the dataset will be 

refined to include further data points, and to incorporate additional descriptors such as specific metal-support 

interactions or surface chemistry characteristics, to improve the model predictive capabilities. Additionally, 

exploring advanced ML techniques, such as deep learning algorithms may offer opportunities to capture more 

complex relationships within the data and to improve the accuracy of predictions. Furthermore, integrating 

kinetic modeling approaches with ML models could provide a deeper understanding of the reaction mechanisms 

and support the production of more robust results. Lastly, conducting experimental studies to validate the 

predictions of the ML framework under real-world FTS conditions is crucial for its practical application in catalyst 

design and process optimization. 

References 

Alayat, A., Echeverria, E., Sotoudehniakarani, F., McIlroy, D.N., McDonald, A.G., 2019, Alumina Coated Silica 

Nanosprings (Ns) Support Based Cobalt Catalysts for Liquid Hydrocarbon Fuel Production from Syngas. 

Materials, 12, 1810. 

Dorneanu, B., Zhang, S., Ruan, H., Heshmat, M., Chen, R., Vassiliadis, V.S., Arellano-Garcia, H., 2022, Big 

Data and Machine Learning: A Roadmap towards Smart Plants. Frontiers of Engineering Management, 9, 

623-639. 

Grainca, A., Tommasi, M., Di Michele, A., Zanella, E., Tonsi, G., Nocchetti, M., Pirola, C., 2023, Study of 

Different H2/CO2 Ratios as Feed in Fischer-Tropsch Reactor with Iron-Based Nano-Hydrotalcite Catalysts. 

Chemical Engineering Transactions, 105, 223-228. 

Kriventceva, E.V., Gryaznov, K.O., Khatkova, K.Y., Sineva, L.V., Mordkovich, V.Z., 2013, The Influence of 

Zeolite Type in Composite Cobalt Catalyst on the Fischer–Tropsch Synthesis Product Composition. Fine 

Chemical Technologies, 8, 9-16. 

Sineva, L., Mordkovich, V., Asalieva, E., Smirnova, V., 2023, Zeolite-Containing Co Catalysts for Fischer–

Tropsch Synthesis with Tailor-Made Molecular-Weight Distribution of Hydrocarbons. Reactions, 43, 359-380 

Sun, B., Qiao, M., Fan, K., Ulrich, J., Tao, F., 2011, Fischer–Tropsch Synthesis over Molecular Sieve Supported 

Catalysts. ChemCatChem, 3, 542-550. 

Teimouri, Z., Abatzoglou, N., Dalai, A.K., 2021, Kinetics and Selectivity Study of Fischer–Tropsch Synthesis to 

C5+ Hydrocarbons: A Review. Catalysts, 11, 330. 

Teimouri, Z., Abatzoglou, N., Dalai, A.K., 2024, A Novel Machine Learning Framework for Designing High-

Performance Catalysts for Production of Clean Liquid Fuels through Fischer-Tropsch Synthesis. Energy, 

289, 130061. 

Tomacruz, J.G.T., Pilario, K.E.S., Remolona, M.F.M., Padama, A.A.B., Ocon, J.D., 2022, A machine-learning 

accelerated density functional theory (ML-DFT) approach for predicting atomic adsorption energies on 

monometallic transition metal surfaces for electrocatalyst screening. Chemical Engineering Transactions, 

94, 733-738. 

588