Next Article in Journal
Recyclable Platinum Nanocatalyst for Nitroarene Hydrogenation: Gum Acacia Polymer-Stabilized Pt Nanoparticles with TiO2 Support
Previous Article in Journal
Adsorption of Methylene Blue on Metakaolin-Based Geopolymers: A Kinetic and Thermodynamic Investigation
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
ChemEngineering
Volume 9
Issue 4
10.3390/chemengineering9040080
ChemEngineering-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Universal Prediction of CO2 Adsorption on Zeolites Using Machine Learning: A Comparative Analysis with Langmuir Isotherm Models

by
Emrah Kirtil
Department of Chemical Engineering, Gebze Technical University, 41400 Kocaeli, Turkey
ChemEngineering 2025, 9(4), 80; https://doi.org/10.3390/chemengineering9040080
Submission received: 25 June 2025 / Revised: 17 July 2025 / Accepted: 22 July 2025 / Published: 28 July 2025
Browse Figures
Versions Notes

Abstract

The global atmospheric concentration of carbon dioxide (CO2) has exceeded 420 ppm. Adsorption-based carbon capture technologies, offer energy-efficient, sustainable solutions. Relying on classical adsorption models like Langmuir to predict CO2 uptake presents limitations due to the need for case-specific parameter fitting. To address this, the present study introduces a universal machine learning (ML) framework using multiple algorithms—Generalized Linear Model (GLM), Feed-forward Multilayer Perceptron (DL), Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), and Gradient Boosted Trees (GBT)—to reliably predict CO2 adsorption capacities across diverse zeolite structures and conditions. By compiling over 5700 experimentally measured adsorption data points from 71 independent studies, this approach systematically incorporates critical factors including pore size, Si/Al ratio, cation type, temperature, and pressure. Rigorous Cross-Validation confirmed superior performance of the GBT model (R2 = 0.936, RMSE = 0.806 mmol/g), outperforming other ML models and providing comparable performance with classical Langmuir model predictions without separate parameter calibration. Feature importance analysis identified pressure, Si/Al ratio, and cation type as dominant influences on adsorption performance. Overall, this ML-driven methodology demonstrates substantial promise for accelerating material discovery, optimization, and practical deployment of zeolite-based CO2 capture technologies.

1. Introduction

The atmospheric concentration of carbon dioxide (CO2) recently surpassed 420 ppm, intensifying worries about global climate change, including sea-level rise, extreme weather events, and ecological impacts [1]. Because fossil fuel use and industrial processes account for roughly 78% of all greenhouse gas emissions, developing effective carbon capture and storage (CCS) technologies is vital to limit environmental harm and meet sustainability goals [2]. Among the various CCS methods, adsorption-based approaches have gained attention thanks to their lower energy needs, operational simplicity, and ability to be regenerated under moderate conditions—qualities that make them appealing alternatives to amine-based absorption [3,4].
Zeolites—microporous aluminosilicates featuring unique pore structures, high surface areas (500–900 m2/g), adjustable Si/Al ratios, and cation-exchange capabilities—are considered promising adsorbents for CO2 capture due to their structural and chemical versatility [5,6]. In these materials, SiO4 and AlO4 tetrahedra form frameworks with microporous channels, where aluminum substitution creates negatively charged sites balanced by extra-framework cations (Na+, K+, Ca2+). These cations strongly interact with CO2 molecules and enhance adsorption [6,7]. For example, zeolite 13X (Si/Al ≈ 1.25, pore size ~0.74 nm) can reach CO2 capacities of about 4.5–4.7 mmol/g under ambient conditions, attributed to its well-matched pore size and strong cation–CO2 interactions [8,9]. Similarly, zeolite 5A (Si/Al ≈ 1.0, pore size ~0.5 nm) achieves capacities around 3.3 mmol/g at 298 K, mainly driven by calcium cation exchange [10].
Historically, adsorption isotherms like the Langmuir model have been used to describe adsorption, but they rely on assumptions of uniform adsorption sites, monolayer coverage, and minimal adsorbate-adsorbate interactions—conditions that often do not hold for real zeolite surfaces [11,12]. In addition, each unique combination of adsorbent and operating conditions (temperature, pressure, humidity) requires its own parameter fitting, which limits the model’s broader applicability [13]. As a result, screening a wide range of adsorbents under different conditions can become time-consuming and less reliable [12,14].
Machine learning (ML) offers a more flexible path toward accurately predicting adsorption capacities. Algorithms like Gradient Boosted Trees, Decision Trees, and Random Forests can handle large, diverse datasets, capturing complex and nonlinear patterns. For instance, Namdeo et al. (2023) employed multiple ML models—including Gradient Boosted Decision Trees—to estimate CO2 uptake in porous carbons, demonstrating improved accuracy over traditional models and identifying key features through SHAP analysis [15]. Similarly, Li et al. (2023) used Random Forests to model CO2 adsorption in MOFs, achieving an R2 of 0.896 and highlighting pore structure and metal centers as critical factors [16]. More recently, Petković et al. (2024) implemented graph neural networks to predict CO2 adsorption in aluminum-substituted zeolites, underscoring the importance of atomistic descriptors and framework topology, while delivering predictions 104–105 times faster than traditional simulation methods [17]. Similarly, Wu et al. (2024) also developed an ML model based on Decision Trees to estimate CO2 uptake in synthetic zeolites from waste-derived materials, achieving strong predictive accuracy with an R2 of approximately 0.87 (RMSE ≈ 0.12 mmol·g−1) [18]. These studies show ML’s potential for building robust, generalized models able to predict adsorption behavior across many materials and conditions.
Despite these advances, ML-based investigations dedicated to zeolite adsorption remain scarce, particularly those leveraging unified datasets that integrate both experimental and computational data. Developing robust predictive models explicitly tailored to zeolite characteristics—such as adsorbent type, pore size, Si/Al ratio, surface area, and cation composition—under various operational parameters (temperature, pressure) would transform how we screen, select, and optimize adsorbents for CCS applications. This study addresses this gap by constructing comprehensive ML models to predict CO2 adsorption capacities of zeolites using an extensive dataset of over 5700 entries drawn from experiments and simulations. We systematically assess the performance of multiple ML algorithms and benchmark their accuracy against classical Langmuir isotherm fittings. By capturing a broad range of physicochemical and operational parameters, these unified models seek to (i) provide reliable, generalizable predictions for diverse zeolite materials and conditions; (ii) identify the most influential factors governing adsorption performance; and (iii) facilitate targeted material design and accelerated CCS development. Ultimately, this work aims to advance CCS research, offering a versatile and practical predictive framework to expedite the discovery and optimization of zeolite-based adsorbents for CO2 capture.

2. Materials and Methods

2.1. Data Collection and Compilation

Adsorption equilibrium data were obtained from the National Institute of Standards and Technology (NIST) online database, specifically targeting experimental studies reporting carbon dioxide (CO2) adsorption on zeolites. Modeling or simulation-based data were excluded to ensure experimental validity for subsequent machine learning analysis. Studies were randomly selected without publication-year restrictions (between the years 1986 and 2021), though each was carefully screened to exclude those lacking sufficient methodological rigor or clearly reported experimental details.
Selected data were downloaded as comma-separated values (.csv) files. An example dataset illustrating the general structure is provided in Supplementary Information (Online Resource S1). These datasets were then compiled into a structured Excel workbook, each assigned to individual sheets, initially totaling 366 separate sheets (Supplementary File, Online Resource S2). Following rigorous manual evaluation to remove anomalous or inconsistent data points, the dataset was refined from 366 to 315 datasets from 71 studies, containing 5765 individual adsorption data points.
Additionally, the original publications of these 71 studies were reviewed to extract relevant zeolite adsorbent parameters not directly included in the adsorption data. Parameters such as zeolite type, Si/Al ratio, pore size, surface area, cation type, temperature, and pressure were documented. Many studies provided multiple adsorbent types or experimental conditions, enhancing the dataset’s variability. All extracted information was consolidated into a unified master Excel spreadsheet, facilitating subsequent analyses. Table 1 provides a detailed summary of references, datasets, and data points for transparency and reproducibility.

2.2. Zeolite Adsorbent Properties

Detailed physicochemical and experimental parameters were systematically extracted from each publication, were subjected to a dimension conversion when necessary, and were incorporated into the unified database to accurately characterize the zeolite adsorbents. The properties collected for analysis and used as input variables in the subsequent machine learning modeling were as follows:
  • Adsorbent type: Zeolite classification based on structural frameworks (e.g., FAU, CHA, ZSM-5, Mordenite, 13X, 5A, 4A).
  • Si/Al molar ratio: The silicon-to-aluminum ratio of zeolite frameworks, critically affecting zeolite hydrophobicity, acidity, and adsorption affinity.
  • Pore size (nm): Average micropore dimensions, directly influencing molecule adsorption capacity and selectivity.
  • Surface area (m2/g): The specific surface area of the adsorbent determined via BET (Brunauer–Emmett–Teller) or comparable nitrogen adsorption methods.
  • Cation type: Predominant exchangeable cation present in zeolite frameworks (e.g., Na+, Ca2+, K+, Li+, NH4+), significantly influencing adsorption interactions through electrostatic attraction and polarization effects.
  • Temperature (°C): Experimental temperature conditions under which adsorption equilibria were recorded, directly impacting adsorption thermodynamics.
  • Pressure (bar): Equilibrium pressure range reported for CO2 adsorption experiments, essential for characterizing adsorption capacity under practical operating conditions.
  • Adsorption capacity (mmol/g): The amount of CO2 adsorbed per unit mass of zeolite under specified conditions, expressed in millimoles per gram. This serves as the primary dependent variable in the study and is predicted using machine learning models based on zeolite properties and operational parameters.
To illustrate the dataset’s comprehensive nature and variability, Figure 1 presents histograms displaying the distribution of each of these critical zeolite properties, clearly depicting their wide variation and ensuring a broad representation of adsorbent properties. This detailed characterization enhances the generalizability and robustness of the machine learning models developed in this study, facilitating predictions across a diverse spectrum of zeolitic materials and operating conditions. Additionally, the final compiled dataset used in this study is provided as a Supplementary File (Online Resource S3) and shared to promote data transparency and facilitate its use by other researchers for further modeling studies.

2.3. Data Preprocessing and Outlier Detection

Given the substantial heterogeneity inherent within the compiled dataset—originating from diverse experimental conditions, zeolite types, structural parameters, and covering a broad publication range—an effective outlier detection step was essential to ensure data quality and reliability. Such variability across the dataset introduced notable inconsistencies and potential inaccuracies, necessitating careful identification and removal of anomalous data points.
Distance-based outlier detection was selected due to its effectiveness in identifying anomalies across multiple parameters simultaneously. In addition to global outliers, local outliers—data points that deviated significantly from nearby values within specific regions of the parameter space—were also identified and removed, as they posed a risk of introducing localized noise into the model. Multiple trials were conducted to determine an optimal threshold, and an 8% threshold was ultimately selected based on these iterative evaluations. Thresholds above 8% did not offer meaningful improvements in model accuracy while reducing the dataset’s representativeness and generalizability to broader conditions. Conversely, thresholds below 8% were insufficient in mitigating noise and inconsistencies. Applying this finalized threshold resulted in the removal of approximately 461 data points from the original 5765, yielding a final dataset of 5304 data points. This carefully optimized preprocessing step significantly enhanced the dataset’s consistency and reliability. The outlier-free version of the dataset is also provided as a Supplementary File (Online Resource S4) to support transparency and reproducibility. Thus, this carefully optimized preprocessing step ensured robust predictive performance without sacrificing the dataset’s integrity or its applicability across diverse zeolite types and experimental conditions.
A missing data analysis revealed that surface area values were unavailable for approximately 32% of the entries in the full dataset (Online Resource S3). All other parameters—Si/Al ratio, pore size, cation type, temperature, and pressure—were complete, with no missing values across the dataset. The missing surface area values resulted from incomplete reporting in some original studies, and despite a thorough review of the publications and Supplementary Materials, these values could not be recovered. Since surface area is a physically meaningful descriptor of adsorption capacity, but its absence in a significant portion of the data posed a challenge, the decision was made to retain the surface area parameter while treating missing values as a separate categorical input using a placeholder flag during model training.
The dataset showed considerable heterogeneity, as it was collected from multiple independent literature studies that used different materials and experimental conditions. To address heterogeneity in the distribution of input parameters (as observed in Figure 1), all continuous variables were normalized to ensure comparable scale across features. No re-sampling or artificial balancing was applied, in order to preserve the original structure of the literature-derived dataset. Instead, machine learning algorithms known to be robust against non-uniform data distributions—such as Random Forest, Gradient Boosted Trees, and feed-forward multilayer perceptron—were employed. Model training was conducted using ten-fold Cross-Validation with random shuffling, maintaining the overall distribution of parameters across folds.

2.4. Machine Learning Model Development

2.4.1. ML Platform and Algorithms

The development, optimization, and evaluation of the machine learning models were performed using the RapidMiner AI Studio software (version 2024.0.1, Educators Edition). In RapidMiner AI Studio, the machine learning workflows were constructed through a visual, block-based interface that allows users to sequentially connect data processing, model training, validation, and evaluation components. Each model was developed by linking modular operators (e.g., “Normalize”, “Cross-Validation”, “Gradient Boosted Trees”) into a pipeline, enabling a partially automated yet fully transparent modeling process. Multiple machine learning algorithms were systematically evaluated to identify the most robust and accurate predictive models for CO2 adsorption on zeolites. These algorithms included the following:
  • Generalized Linear Model (GLM): Selected for its interpretability and ease of implementation, allowing linear and nonlinear relationships to be examined through link functions.
  • Feed-forward Multilayer Perceptron (DL, Deep Neural Network, implemented via RapidMiner’s ‘Deep Learning’ operator): Evaluated for its capacity to capture complex nonlinear interactions among multiple input parameters, especially useful in datasets with diverse features.
  • Decision Tree (DT): Included due to its interpretability and effectiveness in handling heterogeneous datasets by segmenting data into branches based on feature conditions.
  • Random Forest (RF): Assessed for its ability to reduce overfitting through ensemble methods and to capture complex interactions via multiple Decision Trees.
  • Gradient Boosted Trees (GBT): Chosen for their high predictive accuracy and robustness achieved by sequentially building trees, effectively capturing subtle relationships and patterns.
  • Support Vector Machine (SVM): Investigated due to its strength in handling high-dimensional feature spaces and performing well even when clear linear separation between features is lacking.
Each algorithm’s suitability was carefully assessed through multiple iterative trials, reflecting considerable time investment and detailed analytical effort, to ensure selection of the most effective modeling approach.

2.4.2. Data Partitioning and Training

To rigorously assess and enhance model performance, the data were partitioned using a Cross-Validation approach. Cross-Validation, specifically a 10-fold method, was chosen because of its effectiveness in minimizing bias and variance in the model evaluation process. This approach ensured that each data subset was employed both as training and validation sets, facilitating robust accuracy estimates and preventing overfitting.
Hyperparameter optimization was performed for each evaluated algorithm through systematic tuning procedures such as grid search combined with Cross-Validation. This was a particularly intensive process, requiring multiple iterative runs to identify optimal hyperparameter configurations. Additionally, forward feature selection was implemented during model training to further enhance predictive accuracy. Although forward selection significantly increased computational time and complexity, it systematically identified the most informative subset of features, thereby maximizing the accuracy and reliability of the final predictive models.

2.4.3. Feature Engineering

Extensive feature engineering was performed to enhance the predictive power of the input data. This step involved systematically evaluating and transforming the dataset to identify underlying interactions among input parameters. Multiple transformation techniques—including logarithmic, reciprocal, polynomial, and interaction-based features—were rigorously tested to capture nonlinear relationships between zeolite properties (pore size, Si/Al ratio, surface area), adsorption conditions (temperature, pressure), and cation types. Although extensive feature engineering was tested using AI Studio Rapidminer Feature Engineering tool—evaluating over 4800 feature sets and more than 45,000 models—the resulting models exhibited significantly reduced predictive accuracy. In particular, the performance of the best model, Gradient Boosted Trees, dropped from an R2 of 0.936 (using the original features) to 0.426 after feature engineering. The poor performance observed after applying automated feature engineering likely stems from the introduction of noisy or redundant transformations that diluted the signal captured by the original, domain-informed feature set—suggesting that physically meaningful features remain more effective than blindly generated alternatives. This outcome is consistent with prior studies showing that irrelevant or redundant engineered features can degrade model accuracy and increase overfitting [89]. Moreover, multiple researchers have covered the observation that AutoFE techniques such as evolutionary search or black-box transformations frequently result in feature explosion—where the volume of candidate features overwhelms model capacity and introduces noise. This can hinder generalization and model efficiency, as demonstrated in evolutionary AutoFE frameworks like EAAFE and IIFE, which highlighted computational inefficiencies and signal dilution issues in high-dimensional feature spaces [90,91]. Therefore, the final reported results are based on the original feature space, which yielded significantly better performance and generalizability.

2.4.4. Performance Evaluation

Model performance was quantitatively evaluated using three key metrics:
  • Coefficient of Determination (R2): to measure the proportion of variance in adsorption data explained by the model, assessing overall predictive strength.
  • Root Mean Square Error (RMSE): to quantify the magnitude of prediction error, penalizing larger deviations and thus providing a clear indicator of model accuracy.
  • Mean Absolute Error (MAE): to measure average prediction error magnitude, offering intuitive insight into the practical accuracy of model predictions.

2.5. Langmuir Adsorption Isotherm Fitting

2.5.1. Selection of Datasets for Langmuir Model Comparison

To perform an unbiased comparison of the predictive performance of the machine learning models against classical adsorption modeling approaches, four representative adsorption datasets were selected from four different studies (with very different process parameters) and were explicitly excluded from the ML model training. It should be noted that these four external datasets were selected primarily based on the quality and completeness of reported data rather than statistical representativeness across the full zeolite design space. We acknowledge this as a limitation, and present these datasets as illustrative validation cases rather than comprehensive external testing. Detailed information on the selected studies, including the specific zeolite types, number of data points, and references, is provided in Table 2. Additionally, the four datasets extracted from the studies used in this work—including the fitting dataset—are provided in the Supplementary File (Online Resource S5).

2.5.2. Langmuir Model Fitting Procedure

The Langmuir adsorption isotherm was used as a comparative baseline due to its widespread adoption in adsorption studies and its simplicity in describing monolayer adsorption. The mathematical form of the Langmuir equation used in this study is represented as follows:
q e = q m a x K L P 1 + K L P
where qe represents the equilibrium adsorption capacity (mmol/g), qmax is the maximum adsorption capacity (mmol/g), KL denotes the Langmuir equilibrium constant related to the affinity between adsorbate and adsorbent, and P is the equilibrium pressure (bar).
To fit the Langmuir equation to the experimental adsorption data, a custom Python script was developed utilizing a nonlinear regression approach. The Python code used for the Langmuir fitting is provided as a Supplementary File (Online Resource S6) for reproducibility and transparency.
The quality and accuracy of the Langmuir model fitting were quantitatively assessed using two statistical metrics: the coefficient of determination (R2) and the Root Mean Square Error (RMSE). Additionally, visual comparisons of fitted curves against experimental data were generated to qualitatively evaluate the goodness of fit and identify potential systematic deviations.

2.6. Data Presentation and Software

All data organization, management, and table generation were performed using Microsoft Excel 2021 (Version 16.95.1). Machine learning model development, analysis, and visualization were conducted using RapidMiner AI Studio (Educators Edition, Version 2024.0.1). Langmuir adsorption isotherm fitting and associated visualization tasks were carried out using Python within the IDLE environment (Version 3.13.2), utilizing SciPy (version 1.11.4), NumPy (version 1.26.3), Pandas (version 2.1.4), and Matplotlib (version 3.8.2). The authors used a large language model (ChatGPT 4o, OpenAI) solely for AI-assisted copy editing. This included improvements to grammar, clarity, and language consistency in author-written text. No generative content or autonomous scientific writing was produced by the AI. All scientific content, data interpretation, and conclusions were developed and verified entirely by the authors.

3. Results and Discussion

3.1. Dataset Overview

The final dataset used for model development comprises 5765 experimentally measured CO2 adsorption data points collected from 71 independent studies, organized into 315 distinct datasets. The data spans a wide range of material types and adsorption conditions, capturing substantial variability in adsorbent characteristics and thermodynamic parameters. This includes zeolites with diverse pore structures, Si/Al ratios ranging from highly aluminous to siliceous frameworks, various cation types (e.g., Na+, Ca2+, K+), and broad operational ranges of temperature and pressure. This inherent heterogeneity provides a rich foundation for robust machine learning modeling, as it enables the learning of generalized adsorption trends rather than overfitting to a narrow material subclass. The distribution of key parameters across the dataset is presented in Figure 1.
The distribution of adsorbent types across the dataset is notably heterogeneous, with certain zeolite frameworks dominating the data landscape. As shown in Figure 1A, Zeolite 13X is the most frequently studied material, accounting for the largest fraction of data points (over 2000 entries), reflecting its widespread industrial use and well-known performance in CO2 capture applications. Zeolite 4A and Zeolite ZSM-5 also contribute significantly, each with several hundred data points, while other frameworks such as Zeolite 5A, Y, and 4A show moderate representation. In contrast, several less common structures (e.g., Zeolite KFI, DAY, STT, and ZK-5) are represented by relatively few data points. This skewed distribution highlights both the preferential focus in the literature on industrially relevant zeolites and the resulting bias in data availability. Importantly, to avoid overfitting to a specific framework (e.g., 13X), the study models zeolites based on their structural and physicochemical parameters rather than categorical type labels.
The dataset is heavily dominated by Na+-exchanged zeolites, which account for the majority of data points (Figure 1B), reflecting the prevalence of Zeolite 13X and similar sodium forms in CO2 adsorption studies. NH4+ and Ca2+ forms also appear with moderate frequency, while other cation types (e.g., K+, H+, Cu+, Li+, Sr2+) are far less represented. A small number of entries correspond to mixed or unspecified cationic forms.
The temperature distribution in the dataset predominantly ranges from approximately 275 K to 350 K, with the majority of experimental adsorption data collected around ambient conditions (295–310 K), as illustrated in Figure 1C. This reflects typical industrially relevant CO2 adsorption scenarios, notably post-combustion capture processes performed near ambient temperatures. However, the dataset also includes several higher-temperature measurements (above 375 K), broadening the model’s applicability.
The Si/Al ratio distribution (Figure 1D) is highly skewed, with the majority of samples clustered at low ratios between 1 and 5, indicating a high aluminum content and thus a more ionic, cation-rich framework. This is consistent with the dominance of low-silica zeolites such as 13X and 4A in the dataset. A secondary cluster appears around Si/Al ≈ 27, representing more siliceous frameworks like ZSM-5. A limited number of samples extend to very high ratios (up to ~100), though these are rare. This distribution ensures that the model captures adsorption behavior across both cation-dominant and silica-rich frameworks, which is critical given the strong correlation between Si/Al ratio and adsorption energetics [96].
The pore size distribution (Figure 1E) is centered around a dominant peak at approximately 0.74 nm, which corresponds to zeolites with faujasite-type structures (e.g., Zeolite 13X). A significant number of samples also fall in the 0.3–0.5 nm range, reflecting narrower-pore zeolites such as 4A and 5A. While larger pore sizes are present, they are comparatively rare. This skew toward microporous zeolites with pore sizes below 1 nm aligns with their well-established efficiency for selective CO2 adsorption, where molecular sieving and strong confinement effects enhance uptake [97].
As shown in Figure 1F, the dataset contains a wide distribution of pressure values, with a pronounced peak at low-pressure conditions (below 1 bar), followed by a gradual tapering toward higher pressures up to ~45 bar. This skew reflects the typical structure of experimental adsorption isotherms, where multiple pressure points are collected per isotherm to capture the full adsorption profile of a material. Consequently, pressure is the most densely sampled continuous variable in the dataset, comprising a significantly higher number of individual data points compared to other parameters. The high resolution of pressure data enhances the model’s ability to learn pressure-dependent adsorption behavior with precision.
The distribution of CO2 adsorption capacity (Figure 1G), expressed in mmol/g, spans a broad range from near-zero to values exceeding 8 mmol/g, with the majority of data points concentrated between 0 and 6 mmol/g. A notable peak appears around 0–1 mmol/g, reflecting low-pressure or low-affinity adsorption conditions, while a secondary concentration is observed between 3 and 5 mmol/g, indicating favorable adsorption scenarios. As the sole dependent variable in the machine learning model, adsorption capacity represents the key output target. Its high sampling density—comparable to that of pressure—arises from its origin in isotherm data, enabling the model to learn adsorption behavior across varying physical and operational regimes.
From Figure 1, it is clearly evident that the dataset compiled from literature sources is inherently imbalanced in terms of CO2 adsorption capacities. For instance, approximately 62% of the data points correspond to uptake values below 2 mmol g−1, whereas only 8% exceed 5 mmol g−1. This non-uniform distribution reflects the nature of available experimental studies and was expected. As detailed in methods, to mitigate potential bias during model training, all input descriptors were normalized, and machine learning algorithms that are generally robust to skewed data distributions—such as Random Forest and multilayer perceptron—were employed.
Although performance was not evaluated separately across capacity ranges, the overall predictive accuracy remained high (R2 = 0.936 ± 0.012; RMSE = 0.806 ± 0.055 mmol g−1 for the best model, Gradient Boosted Trees), indicating that the observed imbalance did not critically impair model reliability. While this does not compromise the validity of conclusions within the observed data range, predictions made beyond that range should be interpreted with caution owing to increased uncertainty.

3.2. Machine Learning Model Performance

This section evaluates the predictive accuracy of six machine learning algorithms—Generalized Linear Model (GLM), Feed-forward Multilayer Perceptron (Deep Learning Neural Network) (DL), Decision Tree (DT), Random Forest (RF), Gradient Boosted Trees (GBT), and Support Vector Machine (SVM)—using an experimental CO2 adsorption dataset. Key performance metrics include the coefficient of determination (R2), Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE), summarized in Table 3. Predicted versus actual plots for each model (Figure 2A–F) depict their respective performance profiles under various operating conditions and framework compositions.
The GLM yielded limited predictive power (R2 = 0.544, RMSE = 1.929, MAE = 1.377), reflecting the inability of linear formulations to encapsulate nonlinear dependencies arising from pore architecture, surface heterogeneity, and physicochemical interactions between CO2 and framework cations. Adsorption is often governed by site-specific binding energies, electrostatic forces, and cooperative interactions across multiple adsorption sites [98]. Consequently, GLM’s one-to-one mapping between predictors and response fails to account for the synergistic effects that arise from variable partial pressures, cation identity, and Si/Al ratios. As illustrated in Figure 2A, deviations grow with higher CO2 uptakes, highlighting the model’s weakness for capturing the nonlinearity characteristic of adsorption isotherms in porous solids.
The DL model improved upon GLM (R2 = 0.713, RMSE = 1.648, MAE = 1.025; Figure 2B). Neural networks can learn intricate patterns, including the interplay between framework polarity, pore size, and quadrupole interactions from CO2 molecules [99]. However, limited dataset size or noise may precipitate overfitting and local minima convergence, undercutting the advantages of deep architectures. Extensive hyperparameter tuning (e.g., network depth, learning rate, dropout) is typically necessary to prevent overfitting, especially for datasets that span diverse zeolite compositions and pore geometries. Despite these challenges, DL remains a promising approach for adsorption modeling when augmented by larger datasets and optimized regularization strategies [99,100].
Decision Trees produced robust results (R2 = 0.902, RMSE = 0.990, MAE = 0.523; Figure 2C), effectively capturing the nonlinear relationships between adsorbent structure and adsorption capacity via recursive partitioning [101]. Relevant branches often reflect interactions among textural properties (e.g., surface area, pore diameter) and operational variables (e.g., temperature, pressure). However, single-tree models are susceptible to local overfitting, particularly under conditions where certain adsorption regimes dominate the training data. Although visual inspection (Figure 2C) shows potential overestimation at high uptake values, DT remains computationally efficient and interpretable.
Random Forest maintained high accuracy (R2 = 0.909, RMSE = 1.017, MAE = 0.667; Figure 2D). By ensembling multiple Decision Trees, RF mitigates overfitting through bootstrap aggregation, generating robust estimates even when the dataset includes diverse adsorption sites and cationic environments. This ensemble method integrates weaker learners that independently capture complex phase equilibria, cation–CO2 interactions, and pore-diffusion limitations, subsequently averaging their predictions [100,102]. The approach is particularly suited to adsorption problems where heterogeneity of adsorbents and wide parameter variability can produce outliers and localized behaviors.
Gradient Boosted Trees achieved the highest performance (R2 = 0.936, RMSE = 0.806, MAE = 0.458; Figure 2E). GBT’s iterative correction of residuals across sequential trees adeptly captures subtle dependencies—such as synergy among pore topology, cation distribution, and CO2 partial pressure [103]. Internal regularization mechanisms (e.g., learning rate, tree depth constraints) inherently control overfitting, thereby enabling stable generalization. The model’s capacity to resolve hierarchical and second-order interactions underscores its particular effectiveness for adsorption processes spanning broad chemical, structural, and operational domains [15].
SVM delivered intermediate results (R2 = 0.850, RMSE = 1.244, MAE = 0.641; Figure 2F). Its margin-maximizing framework can provide robust predictions for high-dimensional and nonlinear datasets [104]. Nevertheless, kernel selection (e.g., radial basis vs. polynomial) and optimal hyperparameter tuning (e.g., regularization constant, kernel width) are critical to extracting maximum performance. Dataset noise, overlapping feature spaces, and potential parameter misalignment likely constrained SVM’s accuracy. When carefully optimized, SVMs can perform competitively, particularly for moderate-scale adsorption data covering both microporous and mesoporous frameworks.
Aggregating the R2, RMSE, and MAE metrics identifies the top three models as follows:
  • Gradient Boosted Trees (GBT): Highest overall accuracy and robust generalization.
  • Random Forest (RF): Competitive performance with strong resilience to overfitting.
  • Decision Tree (DT): Substantial accuracy coupled with simplicity and fast training.
GBT’s superior performance stems from its iterative boosting strategy and effective regularization, allowing it to disentangle the multilayered interplay of pore geometry, cation identity, and adsorption kinetics. This is consistent with prior findings in CO2 adsorption modeling, where boosting techniques surpass linear and single-tree counterparts [105]. RF and DT remain valuable for rapid prototyping and model interpretability, while DL and SVM warrant exploration with larger datasets and advanced hyperparameter tuning protocols to unlock their full potential.

3.3. Feature Importance and Interpretation

Zeolite adsorption behavior depends on a mix of structural parameters (pore size, Si/Al ratio, surface area, cation type) and operating conditions (pressure, temperature). While a straightforward correlation analysis offers an initial view of how these factors affect CO2 uptake, a more advanced approach—such as Gradient Boosted Trees (GBT)—can uncover nonlinear interactions and better explain how different features work together. Below, we compare results from both methods and discuss how each factor shapes adsorption capacity.

3.3.1. Correlation-Based Analysis

Figure 3 presents the correlation coefficients between individual input parameters and CO2 adsorption capacity. The correlation study showed that temperature (0.306), pressure (0.257), adsorbent type (0.255), and cation type (0.250) have the strongest linear correlations with CO2 uptake. By contrast, surface area (0.066), pore size (0.062), and Si/Al ratio (0.047) appear weaker in simple one-to-one comparisons. These findings generally reflect known thermodynamic principles: higher pressures increase the driving force for gas molecules to adsorb, while higher temperatures lower adsorption capacity by promoting desorption [106]. Zeolite and cation types also influence pore architecture and electrostatic interactions, affecting both selectivity and uptake [107]. However, features like Si/Al ratio and pore size can be underestimated by linear correlation alone because their combined or nonlinear effects may not be evident in simple pairwise metrics [108].

3.3.2. Gradient Boosted Trees (GBT) Model Feature Importance

While correlations provide a useful starting point, the GBT model (R2 = 0.936) offers a deeper perspective on the interplay among parameters. Figure 3 displays the feature importance values obtained from the GBT model, highlighting the ranked contribution of each input parameter to the model’s predictive accuracy. The analysis indicates that pressure (0.474) is the dominant feature, reinforcing the idea that increasing partial pressure strongly promotes CO2 uptake [109]. The Si/Al ratio (0.195) emerges as the second most influential factor, revealing how structural chemistry and framework charge density drive electrostatic interactions with CO2 [96]. Cation type (0.102) also proves critical, since alkali and alkaline-earth metals differ in ionic radius and polarizing power, leading to variations in adsorption affinity [7,110]. Although pore size (0.071) ranks lower, it remains important because microporous channels can either facilitate or hinder diffusion [111]. Surface area (0.054) has long been associated with the sheer number of adsorption sites, but the GBT results suggest that pore connectivity and cation distribution may matter more than total surface area [5]. Adsorbent type (0.021) exerts a smaller direct impact, possibly because its effect on pore structure is partly captured by parameters like pore size and cation identity. Finally, temperature (0.003) shows minimal direct prominence, likely because the model accounts for it indirectly through features such as pressure and framework composition, even though temperature inherently influences desorption [112].
Comparing correlation-based and GBT-based findings highlights the value of comprehensive, multivariate modeling. Linear correlations identify temperature and pressure as primary drivers, whereas the GBT model reveals that parameters like Si/Al ratio, cation type, and pore size carry added weight once nonlinear effects are considered. Although temperature shows a relatively high linear correlation with CO2 uptake, its importance is significantly reduced in the GBT model. This suggests that temperature-dependent effects are likely captured indirectly through pressure and compositional variables, which often co-vary with temperature in the experimental datasets. Such redundancy is resolved automatically in tree-based models that prioritize the most informative split variables. In practice, an elevated temperature may reduce adsorption when viewed in isolation, but its influence can be mitigated by variations in pressure or the type and arrangement of cations when all features are evaluated together. These results underscore the need to consider both thermodynamic conditions and structural attributes for a realistic understanding of zeolite-based CO2 adsorption. Phenomena such as cation gating, framework “breathing,” and the joint roles of pore geometry and electrostatic sites become clearer in advanced models [110,112]. Such insights guide more effective material optimization by indicating which variables—namely pressure, Si/Al ratio, pore size, and cation choice—should be prioritized to achieve higher CO2 capture performance.

3.4. Langmuir Isotherm Modeling and Comparison

The Langmuir isotherm, originally proposed by Langmuir (1918) [113], remains a foundational model for describing gas adsorption on porous materials such as zeolites. The Langmuir isotherm reflects the assumption of a monolayer adsorption process on a homogeneous surface with energetically equivalent sites, where no interactions occur between adsorbed molecules—making it ideal for describing physisorption at low to moderate pressures in microporous solids like zeolites [113,114]. Thanks to its relative simplicity and robust theoretical grounding, it provides critical parameters—such as maximum adsorption capacity and affinity constants—that elucidate adsorbent–adsorbate interactions [11,114].
In microporous zeolites, electrostatic forces from framework cations and the quadrupole moment of CO2 often justify the monolayer assumption at moderate pressures [7]. This solid theoretical basis makes the model especially valuable for describing initial adsorption behavior, and underscores why it remains widely applied in both research and industrial contexts.
Figure 4A–D compare Langmuir adsorption isotherm fits with Gradient Boosted Trees (GBT) predictions for zeolites 13X, 5A, 4A, and Y. Table 4 shows the parameters used in each study in this section. The Langmuir model consistently shows strong alignment with experimental data: R2 values range from 0.9699 (zeolite 5A) to 0.9985 (zeolite Y), while RMSE spans 0.0388 mmol/g (Y) to 0.1817 mmol/g (4A). These metrics reaffirm Langmuir’s suitability for describing monolayer-based uptake in conventional zeolite frameworks [11,114], where uniform energy sites and electrostatic interactions within micropores tend to dominate. In this context, the model’s well-matched performance confirms that a single adsorbate layer often captures the early stages of CO2 adsorption with impressive fidelity.
In contrast, the GBT model—an ensemble of Decision Trees adept at capturing nonlinear interactions—acts as a universal prediction tool that does not require fitting separate parameters for each zeolite or test condition. This is a notable achievement because conventional models like Langmuir typically demand extensive parameter fitting for every new dataset. This asymmetry inherently favors the classical model in these comparisons. However, the ability of the ML model to deliver competitive predictions across systems without individual re-fitting highlights its potential as a scalable alternative when diverse experimental data are available. For zeolites 5A and 4A, the GBT model exhibits high predictive accuracy (R2 of 0.9438 and 0.9857, respectively), sometimes matching or exceeding Langmuir performance (RMSE values of 0.1537 mmol/g and 0.1337 mmol/g). Such robust predictive power underscores how a well-trained machine learning approach can incorporate diverse structural attributes—such as pore size distributions, cation content, and specific textural features—that are often difficult to describe fully using purely mechanistic frameworks. However, the notably lower accuracy for zeolite Y (R2 = 0.7978; RMSE = 0.4558 mmol/g) suggests that certain complexities, including heterogeneous site energies or unusual cation configurations, are not yet fully reflected in the current ML feature set [7,95]. Meanwhile, for zeolite 13X, the GBT model (R2 = 0.9052; RMSE = 0.2591 mmol/g) remains solid but does not outperform Langmuir, demonstrating that a theoretically apt model can still surpass advanced ML under conditions that align closely with monolayer assumptions. Crucially, though, the fact that the ML model can approach Langmuir-level accuracy in many cases, while functioning as a single framework applicable to multiple zeolites, is a significant scientific feat with clear practical benefits.
Such generality is especially advantageous for large-scale screening, design optimization, and operational analyses [17,18]. Once trained on a sufficiently varied dataset, the GBT model rapidly predicts adsorption capacities across an array of zeolite structures and conditions, eliminating the repeated parameter fitting required by classical isotherms. This universal applicability—encompassing different frameworks, Si/Al ratios, and cation compositions—translates into notable time and cost efficiencies for both research and industry, where extensive experimental campaigns are often resource-intensive.
Overall, these results highlight the enduring value of the Langmuir isotherm, particularly at moderate pressures where monolayer coverage remains a valid approximation [115]. Simultaneously, the strong performance of the GBT model for certain zeolites underscores the promise of machine learning in capturing complexities not fully addressed by monolayer-based theories. The flexibility of a universal ML model—a single trained algorithm that can handle different materials and operational conditions—represents a major advance, accelerating material discovery and process development. Future improvements may include the incorporation of more detailed descriptors (e.g., site-specific energy distributions or nuanced cation arrangements) to address limitations such as those observed for zeolite Y. Combining machine learning with established adsorption isotherms could further enhance predictive accuracy and interpretability, merging mechanistic precision with computational efficiency [105]. This synergy of data-driven and theoretical approaches will be key to driving innovation in carbon capture technologies, where precise, reliable adsorption predictions are paramount.

4. Conclusions

This study shows that machine learning models, especially Gradient Boosted Trees (GBT), can reliably predict CO2 adsorption capacities across a wide range of zeolite frameworks. By assembling over 5700 data points from 71 varied sources, we trained and compared multiple algorithms, leading to a universal model that accounts for both the structural complexity of zeolitic materials (pore size, Si/Al ratio, cation type) and important operating conditions (temperature, pressure). Unlike classical Langmuir isotherms, which need separate parameter fitting for each adsorbent or scenario, the GBT model demonstrated notably high generalizability. It often matched or surpassed the accuracy of Langmuir fittings in fresh validation sets, highlighting its usefulness for large-scale screening and optimization of CO2 capture processes.
On a broader level, these outcomes underscore the transformative impact that machine learning can have on carbon capture and storage (CCS). By rapidly screening a wide pool of candidate materials, researchers and engineers can identify the most promising zeolites more efficiently than trial-and-error methods would allow. This knowledge is not only valuable for immediate industrial separation tasks but also points toward the development of “smart” adsorbent platforms, in which data-driven insights guide the deliberate design of next-generation materials. With continued progress in data collection and descriptor quality, machine learning frameworks like this will likely become key tools for advancing adsorption technology and improving its practicality and cost-effectiveness in real-world applications.

Supplementary Materials

The following supporting information can be downloaded at https://www.mdpi.com/article/10.3390/chemengineering9040080/s1, The following Supplementary Materials are provided to support the transparency, reproducibility, and reuse of the data and methods used in this study: Online Resource S1—A sample dataset in .csv format illustrating the raw structure of CO2 adsorption isotherms retrieved from the NIST adsorption database. This file serves as a reference for the formatting and metadata of individual entries. Online Resource S2—A structured .xlsx Excel workbook compiling 366 experimental CO2 adsorption isotherms from literature and NIST, organized with one isotherm per sheet. This unprocessed dataset served as the foundation for further curation and analysis. Online Resource S3—An .xlsx Excel file containing the full machine learning dataset prior to outlier removal, comprising 5765 adsorption data points from 71 experimental studies. Online Resource S4—A cleaned and outlier-filtered version of the ML dataset in .xlsx format, including 5304 high-quality data points used for model training and evaluation. Online Resource S5—An .xlsx Excel file containing external validation data, including experimental isotherms used for performance comparison between ML predictions and Langmuir isotherm fitting. Online Resource S6—A Python script (.py format) implementing nonlinear Langmuir isotherm fitting, performance metric calculations (e.g., RMSE), and comparative visualization.

Funding

This work was supported by the Scientific and Technological Research Council of Türkiye (TÜBİTAK) under the 2219—International Postdoctoral Research Fellowship Programme (Application No: 1059B192302418).

Institutional Review Board Statement

This article does not contain any studies involving human participants or animals performed by any of the authors.

Data Availability Statement

All data supporting the findings of this study are provided as a Supplementary File accompanying this manuscript.

Conflicts of Interest

The authors have no relevant financial or non-financial interests to disclose.

References

  1. Masson-Delmotte, V.; Zhai, P.; Chen, Y.; Goldfarb, L.; Gomis, M.I.; Matthews, J.B.R.; Berger, S.; Huang, M.; Yelekçi, O.; Yu, R.; et al. Climate Change 2021 The Physical Science Basis; Cambridge University Press: Cambridge, UK, 2021; Available online: www.ipcc.ch (accessed on 22 March 2025).
  2. International Energy Agency (IEA). Energy Technology Perspectives: Special Report on Carbon Capture Utilisation and Storage. 2020. Available online: https://www.iea.org/ (accessed on 18 March 2025).
  3. Abanades, J.C.; Arias, B.; Lyngfelt, A.; Mattisson, T.; Wiley, D.E.; Li, H.; Ho, M.T.; Mangano, E.; Brandani, S. Emerging CO2 capture systems. Int. J. Greenh. Gas Control 2015, 40, 126–166. [Google Scholar] [CrossRef]
  4. Choi, S.; Drese, J.H.; Jones, C.W. Adsorbent Materials for Carbon Dioxide Capture from Large Anthropogenic Point Sources. ChemSusChem 2009, 2, 796–854. [Google Scholar] [CrossRef]
  5. Bhavani, A.G.; Reddy, N.S.; Joshi, B.; Sharma, P.; Yadav, P. Enhancing the Adsorption Capacity of CO2 over Modified Microporous Nano-crystalline Zeolite Structure. J. Sci. Res. 2020, 64, 208–211. [Google Scholar] [CrossRef]
  6. Bonenfant, D.; Kharoune, M.; Niquette, P.; Mimeault, M.; Hausler, R. Advances in principal factors influencing carbon dioxide adsorption on zeolites. Sci. Technol. Adv. Mater. 2008, 9, 013007. [Google Scholar] [CrossRef]
  7. Pham, T.D.; Hudson, M.R.; Brown, C.M.; Lobo, R.F. Molecular Basis for the High CO2 Adsorption Capacity of Chabazite Zeolites. ChemSusChem 2014, 7, 3031–3038. [Google Scholar] [CrossRef] [PubMed]
  8. Wang, L.; Liu, Z.; Li, P.; Yu, J.; Rodrigues, A.E. Experimental and modeling investigation on post-combustion carbon dioxide capture using zeolite 13X-APG by hybrid VTSA process. Chem. Eng. J. 2012, 197, 151–161. [Google Scholar] [CrossRef]
  9. Liu, Z.; Shen, C.; Lopes, F.V.S.; Li, P.; Yu, J.; Grande, C.A.; Rodrigues, A.E. Zeolite Apgiia for Adsorption Based Carbon Dioxide Capture. Sep. Sci. Technol. 2013, 48, 388–402. [Google Scholar] [CrossRef]
  10. Yan, P.; Peng, H.; Wu, X.; Rabiee, H.; Weng, Y.; Konarova, M.; Vogrin, J.; Rozhkovskaya, A.; Zhu, Z. Impact of varied zeolite materials on nickel catalysts in CO2 methanation. J. Catal. 2024, 432, 115439. [Google Scholar] [CrossRef]
  11. Sircar, S. Basic Research Needs for Design of Adsorptive Gas Separation Processes. Ind. Eng. Chem. Res. 2006, 45, 5435–5448. [Google Scholar] [CrossRef]
  12. Romá, F.; Riccardo, J.L.; Ramirez-Pastor, A.J. Statistical Thermodynamics Models for Polyatomic Adsorbates: Application to Adsorption of n-Paraffins in 5A Zeolite. Langmuir ACS J. Surf. Colloids 2005, 21, 2454–2459. [Google Scholar] [CrossRef]
  13. Peng, M.; Nguyen, A.V.; Wang, J.; Miller, R. A critical review of the model fitting quality and parameter stability of equilibrium adsorption models. Adv. Colloid Interface Sci. 2018, 262, 50–68. [Google Scholar] [CrossRef]
  14. Zouaoui, E.H.; Djamel, N.; Wan Ba Karim Ghani, W.A.; Amokarane, S. High pressure CO2 adsorption onto NaX zeolite: Effect of Li+, K+, Mg2+ and Zn2+ and equilibrium isotherms study. Iran. J. Chem. Chem. Eng. Int. Engl. Ed. 2021, 40, 1195–1215. [Google Scholar] [CrossRef]
  15. Namdeo, S.; Srivastava, V.C.; Mohanty, P. Machine learning implemented exploration of the adsorption mechanism of carbon dioxide onto porous carbons. J. Colloid Interface Sci. 2023, 647, 174–187. [Google Scholar] [CrossRef]
  16. Li, X.; Zhang, X.; Zhang, J.; Gu, J.; Zhang, S.; Li, G.; Shao, J.; He, Y.; Yang, H.; Zhang, S.; et al. Applied machine learning to analyze and predict CO2 adsorption behavior of metal-organic frameworks. Carbon Capture Sci. Technol. 2023, 9, 100146. [Google Scholar] [CrossRef]
  17. Petković, M.; Vicent-Luna, J.M.; Menkovski, V.; Calero, S. Graph Neural Networks for Carbon Dioxide Adsorption Prediction in Aluminum-Substituted Zeolites. ACS Appl. Mater. Interfaces 2024, 16, 56366–56375. [Google Scholar] [CrossRef] [PubMed]
  18. Wu, H.; Wang, X.; Wang, X.; Su, W. Based on machine learning model for prediction of CO2 adsorption of synthetic zeolite in two-step solid waste treatment. Arab. J. Chem. 2024, 17, 105507. [Google Scholar] [CrossRef]
  19. Yoon, J.W.; Jhung, S.H.; Hwang, Y.K.; Humphrey, S.M.; Wood, P.T.; Chang, J. Gas-Sorption Selectivity of CUK-1: A Porous Coordination Solid Made of Cobalt(II) and Pyridine-2,4- Dicarboxylic Acid. Adv. Mater. 2007, 19, 1830–1834. [Google Scholar] [CrossRef]
  20. Sun, W.; Lin, L.; Peng, X.; Smit, B. Computational screening of porous metal-organic frameworks and zeolites for the removal of SO2 and NOx from flue gases. AIChE J. 2014, 60, 2314–2323. [Google Scholar] [CrossRef]
  21. Talu, O.; Zwiebel, I. Multicomponent adsorption equilibria of nonideal mixtures. AIChE J. 1986, 32, 1263–1276. [Google Scholar] [CrossRef]
  22. Siperstein, F.R.; Myers, A.L. Mixed-gas adsorption. AIChE J. 2001, 47, 1141–1159. [Google Scholar] [CrossRef]
  23. Sievers, W.; Mersmann, A. Single and multicomponent adsorption equilibria of carbon dioxide, nitrogen, carbon monoxide and methane in hydrogen purification processes. Chem. Eng. Technol. 1994, 17, 325–337. [Google Scholar] [CrossRef]
  24. Pirngruber, G.D.; Hamon, L.; Bourrelly, S.; Llewellyn, P.L.; Lenoir, E.; Guillerm, V.; Serre, C.; Devic, T. A Method for Screening the Potential of MOFs as CO2 Adsorbents in Pressure Swing Adsorption Processes. ChemSusChem 2012, 5, 762–776. [Google Scholar] [CrossRef] [PubMed]
  25. Su, X.; Tian, P.; Fan, D.; Xia, Q.; Yang, Y.; Xu, S.; Zhang, L.; Zhang, Y.; Wang, D.; Liu, Z. Synthesis of DNL-6 with a High Concentration of Si (4 Al) Environments and its Application in CO2 Separation. ChemSusChem 2013, 6, 911–918. [Google Scholar] [CrossRef] [PubMed]
  26. Abdul Kareem, F.A.; Shariff, A.M.; Ullah, S.; Garg, S.; Dreisbach, F.; Keong, L.K.; Mellon, N. Experimental and Neural Network Modeling of Partial Uptake for a Carbon Dioxide/Methane/Water Ternary Mixture on 13X Zeolite. Energy Technol. 2017, 5, 1373–1391. [Google Scholar] [CrossRef]
  27. Zheng, Y.; Gu, T. Modified van der Waals Equation for the Prediction of Multicomponent Isotherms. J. Colloid Interface Sci. 1998, 206, 457–463. [Google Scholar] [CrossRef]
  28. Rother, J.; Fieback, T. Multicomponent adsorption measurements on activated carbon, zeolite molecular sieve and metal–organic framework. Adsorption 2013, 19, 1065–1074. [Google Scholar] [CrossRef]
  29. Peter, S.A.; Baron, G.V.; Gascon, J.; Kapteijn, F.; Denayer, J.F.M. Dynamic desorption of CO2 and CH4 from amino-MIL-53(Al) adsorbent. Adsorption 2013, 19, 1235–1244. [Google Scholar] [CrossRef]
  30. Möller, A.; Eschrich, R.; Reichenbach, C.; Guderian, J.; Lange, M.; Möllmer, J. Dynamic and equilibrium-based investigations of CO2-removal from CH4-rich gas mixtures on microporous adsorbents. Adsorption 2017, 23, 197–209. [Google Scholar] [CrossRef]
  31. Wynnyk, K.G.; Hojjati, B.; Pirzadeh, P.; Marriott, R.A. High-pressure sour gas adsorption on zeolite 4A. Adsorption 2017, 23, 149–162. [Google Scholar] [CrossRef]
  32. Shimomura, M.; Yoshida, M.; Endo, A. Influence of free-space calibration using He on the measurement of adsorption isotherms. Adsorption 2017, 23, 249–255. [Google Scholar] [CrossRef]
  33. Pourmahdi, Z.; Maghsoudi, H. Adsorption isotherms of carbon dioxide and methane on CHA-type zeolite synthesized in fluoride medium. Adsorption 2017, 23, 799–807. [Google Scholar] [CrossRef]
  34. Maghsoudi, H.; Aidani, A. Experimental adsorption isotherms of CO2 and CH4 on STT zeolite: Comparison with high- and pure-silica zeolites. Adsorption 2017, 23, 963–969. [Google Scholar] [CrossRef]
  35. Wilkins, N.S.; Rajendran, A. Measurement of competitive CO2 and N2 adsorption on Zeolite 13X for post-combustion CO2 capture. Adsorption 2019, 25, 115–133. [Google Scholar] [CrossRef]
  36. Avijegon, G.; Xiao, G.; Li, G.; May, E.F. Binary and ternary adsorption equilibria for CO2/CH4/N2 mixtures on Zeolite 13X beads from 273 to 333 K and pressures to 900 kPa. Adsorption 2018, 24, 381–392. [Google Scholar] [CrossRef]
  37. Nguyen, H.G.T.; Espinal, L.; van Zee, R.D.; Thommes, M.; Toman, B.; Hudson, M.S.L.; Mangano, E.; Brandani, S.; Broom, D.P.; Benham, M.J.; et al. A reference high-pressure CO2 adsorption isotherm for ammonium ZSM-5 zeolite: Results of an interlaboratory study. Adsorption 2018, 24, 531–539. [Google Scholar] [CrossRef]
  38. Hwang, K.-J.; Hwang, M.J.; Balathanigaimani, M.S.; Nwe, K.; Youn, Y.; Choi, W.-S.; Kim, H.-A.; Nah, J.W.; Shim, W.G. Adsorption characteristics of light gases on basalt rock-based zeolite 4A. Adsorption 2019, 25, 833–842. [Google Scholar] [CrossRef]
  39. Chang, C.-K.; Tun, H.; Chen, C.-C. An activity-based formulation for Langmuir adsorption isotherm. Adsorption 2020, 26, 375–386. [Google Scholar] [CrossRef]
  40. Feng, L.; Shen, Y.; Wu, T.; Liu, B.; Zhang, D.; Tang, Z. Adsorption equilibrium isotherms and thermodynamic analysis of CH4, CO2, CO, N2 and H2 on NaY Zeolite. Adsorption 2020, 26, 1101–1111. [Google Scholar] [CrossRef]
  41. Yang, X.; Kleinrahm, R.; McLinden, M.O.; Richter, M. Uncertainty analysis of adsorption measurements using commercial gravimetric sorption analyzers with simultaneous density measurement based on a magnetic-suspension balance. Adsorption 2020, 26, 645–659. [Google Scholar] [CrossRef]
  42. Tun, H.; Chen, C.-C. Isosteric heat of adsorption from thermodynamic Langmuir isotherm. Adsorption 2021, 27, 979–989. [Google Scholar] [CrossRef]
  43. Tynan, K.J.; Tosso, S.; Ebner, A.D.; Ritter, J.A. On the use of single, dual and three process langmuir models for binary gas mixtures that exhibit unique combinations of these processes. Adsorption 2021, 27, 637–658. [Google Scholar] [CrossRef]
  44. Nguyen, H.G.T.; Toman, B.; Martinez, J.C.; Siderius, D.W.; van Zee, R.D. Reference surface excess isotherms for carbon dioxide adsorption on ammonium ZSM-5 at various temperatures. Adsorption 2022, 28, 15–25. [Google Scholar] [CrossRef]
  45. Shigaki, N.; Mogi, Y.; Haraoka, T.; Furuya, E. Measurements and calculations of the equilibrium adsorption amounts of CO2–N2, CO–N2, and CO2–CO mixed gases on 13X zeolite. SN Appl. Sci. 2020, 2, 488. [Google Scholar] [CrossRef]
  46. Yang, B.; Liu, Y.; Li, M. Separation of CO2–N2 using zeolite NaKA with high selectivity. Chin. Chem. Lett. 2016, 27, 933–937. [Google Scholar] [CrossRef]
  47. Deng, H.; Yi, H.; Tang, X.; Yu, Q.; Ning, P.; Yang, L. Adsorption equilibrium for sulfur dioxide, nitric oxide, carbon dioxide, nitrogen on 13X and 5A zeolites. Chem. Eng. J. 2012, 188, 77–85. [Google Scholar] [CrossRef]
  48. Al-Naddaf, Q.; Rownaghi, A.A.; Rezaei, F. Multicomponent adsorptive separation of CO2, CO, CH4, N2, and H2 over core-shell zeolite-5A@MOF-74 composite adsorbents. Chem. Eng. J. 2020, 384, 123251. [Google Scholar] [CrossRef]
  49. Agueda, V.I.; Delgado, J.A.; Uguina, M.A.; Brea, P.; Spjelkavik, A.I.; Blom, R.; Grande, C. Adsorption and diffusion of H2, N2, CO, CH4 and CO2 in UTSA-16 metal-organic framework extrudates. Chem. Eng. Sci. 2015, 124, 159–169. [Google Scholar] [CrossRef]
  50. Barbosa, G.D.; Travalloni, L.; Castier, M.; Tavares, F.W. Extending an equation of state to confined fluids with basis on molecular simulations. Chem. Eng. Sci. 2016, 153, 212–220. [Google Scholar] [CrossRef]
  51. Liang, Z.; Marshall, M.; Chaffee, A.L. Comparison of Cu-BTC and zeolite 13X for adsorbent based CO2 separation. Energy Procedia 2009, 1, 1265–1271. [Google Scholar] [CrossRef]
  52. Schell, J.; Casas, N.; Marx, D.; Blom, R.; Mazzotti, M. Comparison of commercial and new adsorbent materials for pre-combustion CO2 capture by pressure swing adsorption. Energy Procedia 2013, 37, 167–174. [Google Scholar] [CrossRef]
  53. Tlili, N.; Grévillot, G.; Vallières, C. Carbon dioxide capture and recovery by means of TSA and/or VSA. Int. J. Greenh. Gas Control 2009, 3, 519–527. [Google Scholar] [CrossRef]
  54. Dasgupta, S.; Biswas, N.; Aarti, N.; Gode, N.G.; Divekar, S.; Nanoti, A.; Goswami, A.N. CO2 recovery from mixtures with nitrogen in a vacuum swing adsorber using metal organic framework adsorbent: A comparative study. Int. J. Greenh. Gas Control 2012, 7, 225–229. [Google Scholar] [CrossRef]
  55. Maring, B.J.; Webley, P.A. A new simplified pressure/vacuum swing adsorption model for rapid adsorbent screening for CO2 capture applications. Int. J. Greenh. Gas Control 2013, 15, 16–31. [Google Scholar] [CrossRef]
  56. Bao, Z.; Yu, L.; Ren, Q.; Lu, X.; Deng, S. Adsorption of CO2 and CH4 on a magnesium-based metal organic framework. J. Colloid Interface Sci. 2011, 353, 549–556. [Google Scholar] [CrossRef]
  57. Silva, J.A.; Schumann, K.; Rodrigues, A.E. Sorption and kinetics of CO2 and CH4 in binderless beads of 13X zeolite. Microporous Mesoporous Mater. 2012, 158, 219–228. [Google Scholar] [CrossRef]
  58. Raganati, F.; Ammendola, P.; Chirone, R. CO2 capture performances of fine solid sorbents in a sound-assisted fluidized bed. Powder Technol. 2014, 268, 347–356. [Google Scholar] [CrossRef]
  59. Meng, L.-Y.; Park, S.-J. Influence of MgO template on carbon dioxide adsorption of cation exchange resin-based nanoporous carbon. J. Colloid Interface Sci. 2012, 366, 125–129. [Google Scholar] [CrossRef]
  60. Wu, B.; Zhang, X.; Xu, Y.; Bao, D.; Zhang, S. Assessment of the energy consumption of the biogas upgrading process with pressure swing adsorption using novel adsorbents. J. Clean. Prod. 2015, 101, 251–261. [Google Scholar] [CrossRef]
  61. Mendes, P.A.P.; Ribeiro, A.M.; Gleichmann, K.; Ferreira, A.F.P.; Rodrigues, A.E. Separation of CO2/N2 on binderless 5A zeolite. J. CO2 Util. 2017, 20, 224–233. [Google Scholar] [CrossRef]
  62. AbdulKareem, F.A.; Shariff, A.M.; Ullah, S.; See, T.L.; Keong, L.K.; Mellon, N. Adsorption performance of 5A molecular sieve zeolite in water vapor–binary gas environment: Experimental and modeling evaluation. J. Ind. Eng. Chem. 2018, 64, 173–187. [Google Scholar] [CrossRef]
  63. Youn, H.-K.; Kim, J.; Chandrasekar, G.; Jin, H.; Ahn, W.-S. High pressure carbon dioxide adsorption on nanoporous carbons prepared by Zeolite Y templating. Mater. Lett. 2011, 65, 1772–1774. [Google Scholar] [CrossRef]
  64. Harlick, P.J.; Tezel, F.H. An experimental adsorbent screening study for CO2 removal from N2. Microporous Mesoporous Mater. 2004, 76, 71–79. [Google Scholar] [CrossRef]
  65. Himeno, S.; Tomita, T.; Suzuki, K.; Yoshida, S. Characterization and selectivity for methane and carbon dioxide adsorption on the all-silica DD3R zeolite. Microporous Mesoporous Mater. 2007, 98, 62–69. [Google Scholar] [CrossRef]
  66. Choi, J.-S.; Son, W.-J.; Kim, J.; Ahn, W.-S. Metal–organic framework MOF-5 prepared by microwave heating: Factors to be considered. Microporous Mesoporous Mater. 2008, 116, 727–731. [Google Scholar] [CrossRef]
  67. Ghoufi, A.; Gaberova, L.; Rouquerol, J.; Vincent, D.; Llewellyn, P.; Maurin, G. Adsorption of CO2, CH4 and their binary mixture in Faujasite NaY: A combination of molecular simulations with gravimetry–manometry and microcalorimetry measurements. Microporous Mesoporous Mater. 2009, 119, 117–128. [Google Scholar] [CrossRef]
  68. Pini, R. Interpretation of net and excess adsorption isotherms in microporous adsorbents. Microporous Mesoporous Mater. 2014, 187, 40–52. [Google Scholar] [CrossRef]
  69. Mofarahi, M.; Gholipour, F. Gas adsorption separation of CO2/CH4 system using zeolite 5A. Microporous Mesoporous Mater. 2014, 200, 1–10. [Google Scholar] [CrossRef]
  70. Hefti, M.; Marx, D.; Joss, L.; Mazzotti, M. Adsorption equilibrium of binary mixtures of carbon dioxide and nitrogen on zeolites ZSM-5 and 13X. Microporous Mesoporous Mater. 2015, 215, 215–228. [Google Scholar] [CrossRef]
  71. Yan, B.; Yu, S.; Zeng, C.; Yu, L.; Wang, C.; Zhang, L. Binderless zeolite NaX microspheres with enhanced CO2 adsorption selectivity. Microporous Mesoporous Mater. 2019, 278, 267–274. [Google Scholar] [CrossRef]
  72. Khoramzadeh, E.; Mofarahi, M.; Chung, K.; Lee, C.-H. Equilibrium adsorption and kinetic study of CO2 and N2 on synthesized carbon Black–Zeolite composite. Sep. Purif. Technol. 2022, 280, 119917. [Google Scholar] [CrossRef]
  73. Gholipour, F.; Mofarahi, M. Adsorption equilibrium of methane and carbon dioxide on zeolite 13X: Experimental and thermodynamic modeling. J. Supercrit. Fluids 2016, 111, 47–54. [Google Scholar] [CrossRef]
  74. Moret, S.; Poyet, T.; Bigot, D.; Polster, B.; Crispel, S.; Rouquerol, J.; Llewellyn, P. Complementarity of microcalorimetry, manometry and gravimetry in the study of gas adsorption by microporous solids up to 50 bar. In Studies in Surface Science and Catalysis; Elsevier: Amsterdam, The Netherlands, 2002; Volume 144, pp. 723–729. [Google Scholar]
  75. Sievers, W.; Mersmann, A. Prediction of High Pressure Multicomponent Adsorption Equilibria. In Studies in Surface Science and Catalysis; Elsevier: Amsterdam, The Netherlands, 1993; Volume 80, pp. 583–590. [Google Scholar]
  76. Pakseresht, S.; Kazemeini, M.; Akbarnejad, M.M. Equilibrium isotherms for CO, CO2, CH4 and C2H4 on the 5A molecular sieve by a simple volumetric apparatus. In Separation and Purification Technology; Elsevier: Amsterdam, The Netherlands, 2002; Volume 28, Available online: www.elsevier.com (accessed on 21 July 2025).
  77. Morales-Ospino, R.; Santos, V.N.; Lima, A.R.A.; Torres, A.E.B.; Vilarrasa-García, E.; Bastos-Neto, M.; Cavalcante, C.L.; Azevedo, D.C.S.; Marques, C.R.M.; de Aquino, T.F.; et al. Parametric Analysis of a Moving Bed Temperature Swing Adsorption (MBTSA) Process for Postcombustion CO2 Capture. Ind. Eng. Chem. Res. 2021, 60, 10736–10752. [Google Scholar] [CrossRef]
  78. Shirani, B.; Eic, M. Novel Differential Column Method for Measuring Multicomponent Gas Adsorption Isotherms in NaY Zeolite. Ind. Eng. Chem. Res. 2017, 56, 1008–1018. [Google Scholar] [CrossRef]
  79. Nikolaidis, G.N.; Kikkinides, E.S.; Georgiadis, M.C. An Integrated Two-Stage P/VSA Process for Postcombustion CO2 Capture Using Combinations of Adsorbents Zeolite 13X and Mg-MOF-74. Ind. Eng. Chem. Res. 2017, 56, 974–988. [Google Scholar] [CrossRef]
  80. Yavary, M.; Ebrahim, H.A.; Falamaki, C. Competitive Adsorption Equilibrium Isotherms of CO, CO2, CH4, and H2 on Activated Carbon and Zeolite 5A for Hydrogen Purification. J. Chem. Eng. Data 2016, 61, 3420–3427. [Google Scholar] [CrossRef]
  81. Franco, L.F.M.; Economou, I.G.; Castier, M. Statistical Mechanical Model for Adsorption Coupled with SAFT-VR Mie Equation of State. Langmuir 2017, 33, 11291–11298. [Google Scholar] [CrossRef] [PubMed]
  82. Garcés, S.I.; Villarroel-Rocha, J.; Sapag, K.; Korili, S.A.; Gil, A. Comparative Study of the Adsorption Equilibrium of CO2on Microporous Commercial Materials at Low Pressures. Ind. Eng. Chem. Res. 2013, 52, 6785–6793. [Google Scholar] [CrossRef]
  83. Mason, J.A.; McDonald, T.M.; Bae, T.-H.; Bachman, J.E.; Sumida, K.; Dutton, J.J.; Kaye, S.S.; Long, J.R. Application of a High-Throughput Analyzer in Evaluating Solid Adsorbents for Post-Combustion Carbon Capture via Multicomponent Adsorption of CO2, N2, and H2O. J. Am. Chem. Soc. 2015, 137, 4787–4803. [Google Scholar] [CrossRef]
  84. Cavenati, S.; Grande, C.A.; Rodrigues, A.E. Adsorption Equilibrium of Methane, Carbon Dioxide, and Nitrogen on Zeolite 13X at High Pressures. J. Chem. Eng. Data 2004, 49, 1095–1101. [Google Scholar] [CrossRef]
  85. Aprea, P.; Caputo, D.; Gargiulo, N.; Iucolano, F.; Pepe, F. Modeling Carbon Dioxide Adsorption on Microporous Substrates: Comparison between Cu-BTC Metal−Organic Framework and 13X Zeolitic Molecular Sieve. J. Chem. Eng. Data 2010, 55, 3655–3661. [Google Scholar] [CrossRef]
  86. Calleja, G.; Pau, J.; Calles, J.A. Pure and Multicomponent Adsorption Equilibrium of Carbon Dioxide, Ethylene, and Propane on ZSM-5 Zeolites with Different Si/Al Ratios. J. Chem. Eng. Data 1998, 43, 994–1003. [Google Scholar] [CrossRef]
  87. Loera, S.; Llewellyn, P.L.; Lima, E. Na+ Charge Tuning through Encapsulation of Sulfur Chromophores in Zeolite A and the Consequences in Adsorbent Properties. J. Phys. Chem. C 2010, 114, 7880–7887. [Google Scholar] [CrossRef]
  88. Remy, T.; Peter, S.A.; Van Tendeloo, L.; Van der Perre, S.; Lorgouilloux, Y.; Kirschhock, C.E.A.; Baron, G.V.; Denayer, J.F.M. Adsorption and Separation of CO2 on KFI Zeolites: Effect of Cation Type and Si/Al Ratio on Equilibrium and Kinetic Properties. Langmuir 2013, 29, 4998–5012. [Google Scholar] [CrossRef] [PubMed]
  89. Kaur, B. Feature engıneerıng in machıne learnıng: Selectıon, extractıon, and theır ımpact on model performance. Int. J. Inf. Technol. Manag. Inf. Syst. 2025, 16, 1–18. [Google Scholar] [CrossRef]
  90. Overman, T.; Klabjan, D.; Utke, J. IIFE: Interaction Information Based Automated Feature Engineering. In Proceedings of the 2024 IEEE International Conference on Data Mining (ICDM), Abu Dhabi, United Arab Emirates, 9–12 December 2024; pp. 803–808. [Google Scholar] [CrossRef]
  91. Zhu, G.; Jiang, S.; Guo, X.; Yuan, C.; Huang, Y. Evolutionary Automated Feature Engineering. Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 13629 LNCS. In Pacific Rim International Conference on Artificial Intelligence; Springer Nature: Cham, Switzerland, 2022; pp. 574–586. [Google Scholar] [CrossRef]
  92. Liu, J.; Wang, Y.; Benin, A.I.; Jakubczak, P.; Willis, R.R.; LeVan, M.D. CO2/H2O Adsorption Equilibrium and Rates on Metal−Organic Frameworks: HKUST-1 and Ni/DOBDC. Langmuir 2010, 26, 14301–14307. [Google Scholar] [CrossRef]
  93. Liu, Z.; Grande, C.A.; Li, P.; Yu, J.; Rodrigues, A.E. Adsorption and Desorption of Carbon Dioxide and Nitrogen on Zeolite 5A. Sep. Sci. Technol. 2011, 46, 434–451. [Google Scholar] [CrossRef]
  94. Figini-Albisetti, A.; Velasco, L.F.; Parra, J.B.; Ania, C.O. Effect of outgassing temperature on the performance of porous materials. Appl. Surf. Sci. 2010, 256, 5182–5186. [Google Scholar] [CrossRef]
  95. Llewellyn, P.L.; Bourrelly, S.; Vagner, C.; Heymans, N.; Leclerc, H.; Ghoufi, A.; Bazin, P.; Vimont, A.; Daturi, M.; Devic, T.; et al. Evaluation of MIL-47(V) for CO2-Related Applications. J. Phys. Chem. C 2013, 117, 962–970. [Google Scholar] [CrossRef]
  96. Choi, H.J.; Hong, S.B. Effect of framework Si/Al ratio on the mechanism of CO2 adsorption on the small-pore zeolite gismondine. Chem. Eng. J. 2022, 433, 133800. [Google Scholar] [CrossRef]
  97. Tao, Z.; Tian, Y.; Wu, W.; Liu, Z.; Fu, W.; Kung, C.-W.; Shang, J. Development of zeolite adsorbents for CO2 separation in achieving carbon neutrality. NPJ Mater. Sustain. 2024, 2, 20. [Google Scholar] [CrossRef]
  98. Myers, A.L.; Monson, P.A. Adsorption in Porous Materials at High Pressure: Theory and Experiment. Langmuir 2002, 18, 10261–10273. [Google Scholar] [CrossRef]
  99. LeCun, Y.; Bengio, Y.; Hinton, G. Deep learning. Nature 2015, 521, 436–444. [Google Scholar] [CrossRef] [PubMed]
  100. Mahmoud, H.A.H.; Hakami, N.A.; Hafez, A.M.; Kooh, M.R.R. An Intelligent Deep Learning Model for CO2 Adsorption Prediction. Adsorpt. Sci. Technol. 2022, 2022, 8136302. [Google Scholar] [CrossRef]
  101. Breiman, L.; Friedman, J.H.; Olshen, R.A.; Stone, C.J. Classification and Regression Trees; Chapman and Hall/CRC: New York, NY, USA, 2017; pp. 1–358. [Google Scholar] [CrossRef]
  102. Deng, X.; Yang, W.; Li, S.; Liang, H.; Shi, Z.; Qiao, Z. Large-Scale Screening and Machine Learning to Predict the Computation-Ready, Experimental Metal-Organic Frameworks for CO2 Capture from Air. Appl. Sci. 2020, 10, 569. [Google Scholar] [CrossRef]
  103. Friedman, J.H. Greedy function approximation: A gradient boosting machine. Ann. Stat. 2001, 29, 1189–1232. [Google Scholar] [CrossRef]
  104. Cortes, C.; Vapnik, V.; Saitta, L. Support-vector networks. Mach. Learn. 1995, 20, 273–297. [Google Scholar] [CrossRef]
  105. Gu, Y.-T.; Gu, Y.-M.; Tao, Q.; Wang, X.; Zhu, Q.; Ma, J. Machine learning for prediction of CO2/N2/H2O selective adsorption and separation in metal-zeolites. J. Mater. Inform. 2023, 3, 19. [Google Scholar] [CrossRef]
  106. Kim, M.-H.; Cho, I.-H.; Choi, S.-O.; Choo, S.-T. Zeolites: Their Features as Pressure Swing Adsorbents and CO2 Adsorption Capacity. J. Environ. Sci. Int. 2014, 23, 943–962. [Google Scholar] [CrossRef]
  107. Dabbawala, A.A.; Ismail, I.; Vaithilingam, B.V.; Polychronopoulou, K.; Singaravel, G.; Morin, S.; Berthod, M.; Al Wahedi, Y. Synthesis of hierarchical porous Zeolite-Y for enhanced CO2 capture. Microporous Mesoporous Mater. 2020, 303, 110261. [Google Scholar] [CrossRef]
  108. Jhung, S.H.; Yoon, J.W.; Lee, J.S.; Chang, J. Low-Temperature Adsorption/Storage of Hydrogen on FAU, MFI, and MOR Zeolites with Various Si/Al Ratios: Effect of Electrostatic Fields and Pore Structures. Chem. Eur. J. 2007, 13, 6502–6507. [Google Scholar] [CrossRef]
  109. Sayari, A.; Belmabkhout, Y.; Serna-Guerrero, R. Flue gas treatment via CO2 adsorption. Chem. Eng. J. 2011, 171, 760–774. [Google Scholar] [CrossRef]
  110. Choi, H.J.; Bruce, E.L.; Kencana, K.S.; Hong, J.; Wright, P.A.; Hong, S.B. Highly Cooperative CO2 Adsorption via a Cation Crowding Mechanism on a Cesium-Exchanged Phillipsite Zeolite. Angew. Chem. Int. Ed. Engl. 2023, 62, e202305816. [Google Scholar] [CrossRef]
  111. Yang, J.; Li, J.; Wang, W.; Li, L.; Li, J. Adsorption of CO2, CH4, and N2 on 8-, 10-, and 12-Membered Ring Hydrophobic Microporous High-Silica Zeolites: DDR, Silicalite-1, and Beta. Ind. Eng. Chem. Res. 2013, 52, 17856–17864. [Google Scholar] [CrossRef]
  112. Orsikowsky-Sanchez, A.; Plantier, F.; Miqueu, C. Coupled gravimetric, manometric and calorimetric study of CO2, N2 and CH4 adsorption on zeolites for the assessment of classical equilibrium models. Adsorption 2020, 26, 1137–1152. [Google Scholar] [CrossRef]
  113. Langmuir, I. The adsorption of gases on plane surfaces of glass, mica and platinum. J. Am. Chem. Soc. 1918, 40, 1361–1403. [Google Scholar] [CrossRef]
  114. Ruthven, D.M. Dynamics of adsorption columns: Single-transition systems. In Principles of Adsorption and Adsorption Processes; John Wiley & Sons, Inc.: Hoboken, NJ, USA, 1984; p. 433. Available online: https://www.wiley.com/en-us/Principles+of+Adsorption+and+Adsorption+Processes-p-9780471866060 (accessed on 21 March 2025).
  115. Lu, X.; Rothstein, D.; Madey, R.; Huang, J.-C. Pressure Swing Adsorption for a System with a Langmuir Isotherm. Sep. Sci. Technol. 1988, 23, 281–291. [Google Scholar] [CrossRef]
Chemengineering 09 00080 g001aChemengineering 09 00080 g001bChemengineering 09 00080 g001cChemengineering 09 00080 g001d
Figure 1. Distribution of key input and output parameters in the machine learning dataset: (A) adsorbent type, (B) cation type, (C) temperature, (D) Si/Al ratio, (E) pore size, (F) pressure, and (G) CO2 adsorption capacity.
Figure 1. Distribution of key input and output parameters in the machine learning dataset: (A) adsorbent type, (B) cation type, (C) temperature, (D) Si/Al ratio, (E) pore size, (F) pressure, and (G) CO2 adsorption capacity.
Chemengineering 09 00080 g001aChemengineering 09 00080 g001bChemengineering 09 00080 g001cChemengineering 09 00080 g001d
Chemengineering 09 00080 g002aChemengineering 09 00080 g002bChemengineering 09 00080 g002c
Figure 2. Predicted versus actual CO2 adsorption capacities (mmol/g) for six machine learning models: (A) Generalized Linear Model (GLM), (B) Feed-forward Multilayer Perceptron (DL), (C) Decision Tree (DT), (D) Random Forest (RF), (E) Gradient Boosted Trees (GBT), and (F) Support Vector Machine (SVM), (.Chemengineering 09 00080 i001) perfect prediction (y = x), Chemengineering 09 00080 i002: individual data points.
Figure 2. Predicted versus actual CO2 adsorption capacities (mmol/g) for six machine learning models: (A) Generalized Linear Model (GLM), (B) Feed-forward Multilayer Perceptron (DL), (C) Decision Tree (DT), (D) Random Forest (RF), (E) Gradient Boosted Trees (GBT), and (F) Support Vector Machine (SVM), (.Chemengineering 09 00080 i001) perfect prediction (y = x), Chemengineering 09 00080 i002: individual data points.
Chemengineering 09 00080 g002aChemengineering 09 00080 g002bChemengineering 09 00080 g002c
Chemengineering 09 00080 g003
Figure 3. Comparison of feature importance values from correlation analysis and Gradient Boosted Trees (GBT) model.
Figure 3. Comparison of feature importance values from correlation analysis and Gradient Boosted Trees (GBT) model.
Chemengineering 09 00080 g003
Chemengineering 09 00080 g004aChemengineering 09 00080 g004b
Figure 4. Comparison of experimental CO2 adsorption isotherms, Langmuir model fits, and Gradient Boosted Trees (GBT) predictions for four external test datasets from the studies: (A) [92], (B) [93], (C) [94], and (D) [95].
Figure 4. Comparison of experimental CO2 adsorption isotherms, Langmuir model fits, and Gradient Boosted Trees (GBT) predictions for four external test datasets from the studies: (A) [92], (B) [93], (C) [94], and (D) [95].
Chemengineering 09 00080 g004aChemengineering 09 00080 g004b
Table 1. Overview of the studies included in the machine learning dataset.
Table 1. Overview of the studies included in the machine learning dataset.
Study IDNameType of ZeoliteCationNumber of DatasetsNumber of DataReference
1Gas-Sorption Selectivity of CUK-1: A Porous Coordination Solid Made of Cobalt(II) and Pyridine-2,4-Dicarboxylic AcidZeolite 4ANa+19[19]
2Computational Screening of Porous Metal-Organic Frameworks and Zeolites for the Removal of SO2 and NOₓ from Flue GasesZeolite FAUNone16[20]
3Multicomponent Adsorption Equilibria of Nonideal MixturesH-MordeniteH+256[21]
4Mixed-Gas AdsorptionZeolite 13XNa+258[22]
5Single- and Multicomponent Adsorption Equilibria of Carbon Dioxide, Nitrogen, Carbon Monoxide, and Methane in Hydrogen Purification ProcessesZeolite 5ACa2+/Na+119[23]
6A Method for Screening the Potential of MOFs as CO2 Adsorbents in Pressure Swing Adsorption ProcessesZeolite 13XNa+228[24]
7Synthesis of DNL-6 with a High Concentration of Si (4 Al) Environments and its Application in CO2 SeparationZeolite ANa+123[25]
8Experimental and Neural Network Modeling of Partial Uptake for a Carbon Dioxide/Methane/Water Ternary Mixture on 13X ZeoliteZeolite 13XNa+326[26]
9Modified van der Waals Equation for the Prediction of Multicomponent IsothermsZeolite YCu+/Na+682[27]
10Multicomponent Adsorption Measurements on Activated Carbon, Zeolite Molecular Sieve, and Metal–Organic FrameworkZeolite 13XNa+324[28]
11Dynamic Desorption of CO2 and CH4 from Amino-MIL-53(Al) AdsorbentZeolite 13XNa+334[29]
12Dynamic and Equilibrium-Based Investigations of CO2 Removal from CH4-Rich Gas Mixtures on Microporous AdsorbentsZeolite 13XNa+398[30]
13High-Pressure Sour Gas Adsorption on Zeolite 4AZeolite 4ANa+6568[31]
14Influence of Free-Space Calibration Using He on the Measurement of Adsorption IsothermsZeolite YNa+144[32]
15Adsorption Isotherms of Carbon Dioxide and Methane on CHA-Type Zeolite Synthesized in Fluoride MediumZeolite CHAH+/Na+438[33]
16Experimental Adsorption Isotherms of CO2 and CH4 on STT Zeolite: Comparison with High- and Pure-Silica ZeolitesZeolite STTNone678[34]
17Measurement of Competitive CO2 and N2 Adsorption on Zeolite 13X for Post-Combustion CO2 CaptureZeolite 13XNa+6240[35]
18Binary and Ternary Adsorption Equilibria for CO2/CH4/N2 Mixtures on Zeolite 13X Beads from 273 to 333 K and Pressures to 900 kPaZeolite 13XNa+977[36]
19A Reference High-Pressure CO2 Adsorption Isotherm for Ammonium ZSM-5 Zeolite: Results of an Interlaboratory StudyZeolite ZSM-5NH4+13344[37]
20Adsorption Characteristics of Light Gases on Basalt Rock-Based Zeolite 4AZeolite 4ANa+390[38]
21An activity-based formulation for Langmuir adsorption isothermZeolite 5ACa2+339[39]
22Adsorption equilibrium isotherms and thermodynamic analysis of CH4, CO2, CO, N2, and H2 on NaY ZeoliteZeolite YNa+4108[40]
23Uncertainty analysis of adsorption measurements using commercial gravimetric sorption analyzers with simultaneous density measurement based on a magnetic-suspension balanceZeolite 13XNa+143[41]
24Isosteric heat of adsorption from thermodynamic Langmuir isothermZeolite 13XNa+10279[42]
25On the use of single-, dual-, and three-process Langmuir models for binary gas mixtures that exhibit unique combinations of these processesZeolite 13XNa+573[43]
26Reference surface excess isotherms for carbon dioxide adsorption on ammonium ZSM-5 at various temperaturesZeolite ZSM-5NH4+12278[44]
27Measurements and calculations of the equilibrium adsorption amounts of CO2–N2, CO–N2, and CO2–CO mixed gases on 13X zeoliteZeolite 13XNa+684[45]
28Separation of CO2–N2 using zeolite NaKA with high selectivityZeolite ANa+/K+844[46]
29Adsorption equilibrium for sulfur dioxide, nitric oxide, carbon dioxide, nitrogen on 13X and 5A zeolitesZeolite 5ACa2+14164[47]
30Multicomponent adsorptive separation of CO2, CO, CH4, N2, and H2 over core-shell zeolite-5A@MOF-74 composite adsorbentsZeolite 5ACa2+138[48]
31Adsorption and diffusion of H2, N2, CO, CH4, and CO2 in UTSA-16 metal-organic framework extrudatesZeolite 13XNa+142[49]
32Extending an equation of state to confined fluids with basis on molecular simulationsZeolite 13XNa+226[50]
33Comparison of Cu-BTC and zeolite 13X for adsorbent-based CO2 separationZeolite 13XNa+19[51]
34Comparison of commercial and new adsorbent materials for pre-combustion CO2 capture by pressure swing adsorptionZeolite 13XNa+131[52]
35Carbon dioxide capture and recovery by means of TSA and/or VSAZeolite 5ACa2+10100[53]
36CO2 recovery from mixtures with nitrogen in a vacuum swing adsorber using metal-organic framework adsorbent: A comparative studyZeolite 13XNa+221[54]
37A new simplified pressure/vacuum swing adsorption model for rapid adsorbent screening for CO2 capture applicationsZeolite 13XNa+3122[55]
38Adsorption of CO2 and CH4 on a magnesium-based metal-organic frameworkZeolite 13XNa+672[56]
39Sorption and kinetics of CO2 and CH4 in binderless beads of 13X zeoliteZeolite 13XNa+324[57]
40CO2 capture performances of fine solid sorbents in a sound-assisted fluidized bedZeolite 13XNa+12[58]
41Influence of MgO template on carbon dioxide adsorption of cation exchange resin-based nanoporous carbonZeolite 13XNa+114[59]
42Assessment of the energy consumption of the biogas upgrading process with pressure swing adsorption using novel adsorbentsZeolite 13XNa+231[60]
43Separation of CO2/N2 on binderless 5A zeoliteZeolite 5ACa2+236[61]
44Adsorption performance of 5A molecular sieve zeolite in water vapor–binary gas environment: Experimental and modeling evaluationZeolite 5ACa2+/Na+220[62]
45High-pressure carbon dioxide adsorption on nanoporous carbons prepared by Zeolite Y templatingZeolite 13XNa+124[63]
46An experimental adsorbent screening study for CO2 removal from N2Zeolite YNa+868[64]
47Characterization and selectivity for methane and carbon dioxide adsorption on the all-silica DD3R zeoliteZeolite DD3RNone453[65]
48Metal–organic framework MOF-5 prepared by microwave heating: Factors to be consideredZeolite 13XNa+110[66]
49Adsorption of CO2, CH4, and their binary mixture in Faujasite NaY: A combination of molecular simulations with gravimetry–manometry and microcalorimetry measurementsZeolite YNa+8123[67]
50Interpretation of net and excess adsorption isotherms in microporous adsorbentsZeolite 13XNa+294[68]
51Gas adsorption separation of CO2/CH4 system using zeolite 5AZeolite 5ACa2+244[69]
52Adsorption equilibrium of binary mixtures of carbon dioxide and nitrogen on zeolites ZSM-5 and 13XZeolite ZSM-5H+361[70]
53Binderless zeolite NaX microspheres with enhanced CO2 adsorption selectivityZeolite NaX@NaANa+6132[71]
54Equilibrium adsorption and kinetic study of CO2 and N2 on synthesized carbon Black–Zeolite compositeZeolite 13XNa+328[72]
55Adsorption equilibrium of methane and carbon dioxide on zeolite 13X: Experimental and thermodynamic modelingZeolite 13XNa+599[73]
57Complementarity of microcalorimetry, manometry, and gravimetry in the study of gas adsorption by microporous solids up to 50 barZeolite 13XNa+550[74]
58CO2 adsorption in faujasite systems: microcalorimetry and molecular simulationZeolite YNa+499[67]
59Prediction of High-Pressure Multicomponent Adsorption EquilibriaZeolite 5ACa2+/Na+121[75]
60Equilibrium isotherms for CO, CO2, CH4, and C2H4 on the 5A molecular sieve by a simple volumetric apparatusZeolite 5ACa2+/Na+682[76]
61Parametric Analysis of a Moving Bed Temperature Swing Adsorption (MBTSA) Process for Post-combustion CO2 CaptureZeolite 13XNa+569[77]
62Novel Differential Column Method for Measuring Multicomponent Gas Adsorption Isotherms in NaY ZeoliteZeolite YNa+693[78]
63An Integrated Two-Stage P/VSA Process for Post-combustion CO2 Capture Using Combinations of Adsorbents Zeolite 13X and Mg-MOF-74Zeolite 13XNa+257[79]
64Competitive Adsorption Equilibrium Isotherms of CO, CO2, CH4, and H2 on Activated Carbon and Zeolite 5A for Hydrogen PurificationZeolite 5ACa2+/Na+318[80]
65Statistical Mechanical Model for Adsorption Coupled with SAFT-VR Mie Equation of StateZeolite 13XNa+571[81]
66Comparative Study of the Adsorption Equilibrium of CO2 on Microporous Commercial Materials at Low PressuresZeolite 13XNa+353[82]
67Application of a High-Throughput Analyzer in Evaluating Solid Adsorbents for Post-Combustion Carbon Capture via Multicomponent Adsorption of CO2, N2, and H2OZeolite 13XNa+238[83]
68Adsorption Equilibrium of Methane, Carbon Dioxide, and Nitrogen on Zeolite 13X at High PressuresZeolite 13XNa+346[84]
69Modeling Carbon Dioxide Adsorption on Microporous Substrates: Comparison between Cu-BTC Metal-Organic Framework and 13X Zeolitic Molecular SieveZeolite 13XNa+651[85]
71Application of a High-Throughput Analyzer in Evaluating Solid Adsorbents for Post-Combustion Carbon Capture via Multicomponent Adsorption of CO2, N2, and H2OZeolite ZSM-5Na+658[86]
72Adsorption Equilibrium of Methane, Carbon Dioxide, and Nitrogen on Zeolite 13X at High PressuresZeolite 4ANa+6313[87]
73Modeling Carbon Dioxide Adsorption on Microporous Substrates: Comparison between Cu-BTC Metal-Organic Framework and 13X Zeolitic Molecular SieveZeolite KFILi+32268[88]
Table 2. Overview of the four external studies selected for model comparison.
Table 2. Overview of the four external studies selected for model comparison.
Study IDNameType of ZeoliteCationNumber of DataReference
L1CO2/H2O Adsorption Equilibrium and Rates on Metal-Organic Frameworks: HKUST-1 and Ni/DOBDCZeolite 13XNa+15[92]
L2Adsorption and Desorption of Carbon Dioxide and Nitrogen on Zeolite 5AZeolite 5ACa2+11[93]
L3Effect of Outgassing Temperature on the Performance of Porous MaterialsZeolite 4ANa+13[94]
L4Evaluation of MIL-47(V) for CO2-Related ApplicationsZeolite YNa+19[95]
Table 3. Performance comparison of six machine learning models.
Table 3. Performance comparison of six machine learning models.
ModelR2RMSE (mmol/g)MAE (mmol/g)
Generalized Linear Model0.544 ± 0.0341.929 ± 0.1561.377 ± 0.034
Feed-forward Multilayer Perceptron0.713 ± 0.0291.648 ± 0.1521.025 ± 0.026
Decision Tree0.902 ± 0.0230.990 ± 0.1190.523 ± 0.044
Random Forest0.909 ± 0.0091.017 ± 0.0680.667 ± 0.016
Gradient Boosted Trees0.936 ± 0.0120.806 ± 0.0550.458 ± 0.023
Support Vector Machine0.850 ± 0.0321.244 ± 0.1210.641 ± 0.028
Table 4. Zeolite properties and experimental conditions for the external test datasets.
Table 4. Zeolite properties and experimental conditions for the external test datasets.
AuthorsAdsorbent TypeCation TypeTemperature (K)Si Al RatioPore Size (nm)Surface Area (m2/g)
Liu et al. (2010) [92]Zeolite 13XNa+2981.3751590
Liu et al. (2011) [93]Zeolite 5ACa2+30310.5450
Figini-Albisetti et al. (2010) [94]Zeolite 4ANa+29810.41500
Llewellyn et al. (2013) [95]Zeolite YNa+2982.60.74700
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Kirtil, E. Universal Prediction of CO2 Adsorption on Zeolites Using Machine Learning: A Comparative Analysis with Langmuir Isotherm Models. ChemEngineering 2025, 9, 80. https://doi.org/10.3390/chemengineering9040080

AMA Style

Kirtil E. Universal Prediction of CO2 Adsorption on Zeolites Using Machine Learning: A Comparative Analysis with Langmuir Isotherm Models. ChemEngineering. 2025; 9(4):80. https://doi.org/10.3390/chemengineering9040080

Chicago/Turabian Style

Kirtil, Emrah. 2025. "Universal Prediction of CO2 Adsorption on Zeolites Using Machine Learning: A Comparative Analysis with Langmuir Isotherm Models" ChemEngineering 9, no. 4: 80. https://doi.org/10.3390/chemengineering9040080

APA Style

Kirtil, E. (2025). Universal Prediction of CO2 Adsorption on Zeolites Using Machine Learning: A Comparative Analysis with Langmuir Isotherm Models. ChemEngineering, 9(4), 80. https://doi.org/10.3390/chemengineering9040080

Article Metrics

Back to TopTop