Ahmad Mohaddespour

The efficient removal of low-concentration nitrogen (N₂) is a critical and challenging task in the industrial production of high-purity oxygen (O₂), particularly in air separation processes and in natural gas purification for generating high-purity methane (CH₄). In this work, three advanced modeling techniques, namely the group method of data handling (GMDH), gene expression programming (GEP), and genetic programming (GP) were applied to develop user-friendly mathematical correlations for predicting nitrogen storage capacity of MOFs. A broad dataset of 3073 laboratory measurements was employed. The developed models were further evaluated for their accuracy and reliability using various statistical and graphical methods. Among the models, the GEP correlation provided the most reliable outcomes with superior statistical metrics, yielding mean absolute error (MAE) values of 0.9924, 1.0101 and 0.9959 for the training, testing, and overall datasets, and R² values of 0.9703, 0.9750, and 0.9714, respectively. All the models closely followed the expected trend of N₂ storage under varying pressure. Additionally, the Pearson, Spearman, and Kendall correlation analyses were employed to examine the influence of individual input factors on the model outputs. The findings revealed that temperature has the most substantial impact on storage capacity across both linear and non-linear dimensions, whereas pressure primarily affects it through nonlinear interactions. At last, the credibility of the collected databank and the applicability of the recommended correlations were confirmed using the leverage method, demonstrating more than 95% of the collected data fell within the acceptable range of the Williams plot.

Accurate prediction of methane (CH₄) adsorption in coal and shale formations is essential for optimizing gas extraction, enhancing storage efficiency, mitigating greenhouse gas emissions, and ensuring safe mining and unconventional reservoir development. In this study, the most comprehensive dataset reported to date was compiled to model CH₄ adsorption in coal and shale systems, using pressure (P), moisture content (M), and total organic carbon (TOC) as inputs for coal, and temperature (T), moisture content (M), and TOC for shale. To simultaneously achieve high predictive accuracy and physical interpretability, three transparent white-box models gene expression programming (GEP), group method of data handling (GMDH), and genetic programming (GP) and three advanced black-box models Gaussian process regression (GPR), categorical boosting (CatBoost), and light gradient boosting machine (LightGBM) were developed and systematically evaluated. The white-box models successfully produced explicit, closed-form mathematical correlations that directly link CH₄ adsorption to key controlling parameters, providing transparent, physically interpretable tools that can be readily applied without specialized software. In parallel, the black-box models were employed to benchmark predictive performance, with CatBoost achieving the best overall accuracy, yielding mean squared errors (MSE) of 0.008 and 0.004 and coefficients of determination (R²) of 0.924 and 0.997 for coal and shale, respectively. Model interpretability was further enhanced using SHapley Additive exPlanations (SHAP), which consistently identified TOC as the dominant factor governing CH₄ adsorption in both systems, while leverage-based outlier analysis confirmed the robustness of the optimal models, with more than 99% of coal data and over 95% of shale data falling within the applicability domain. Notably, this work represents the first application of GEP, GMDH, and GP to derive transparent, user-friendly analytical expressions for CH₄ adsorption in coal, offering a practical and interpretable alternative to purely black-box machine-learning approaches.

Accurate estimation of carbon dioxide (CO₂) equilibrium solubility in absorbent solutions is essential for designing and optimizing chemical absorption-based CO₂ capture systems. In this study, in order to simulate the CO ₂ loading capacity of Diethanolamine (DEA) aqueous solutions, 3 renowned white box approaches, including the Group Method of Data Handling (GMDH), Genetic Programming (GP), and Gene Expression Programming (GEP), were used. To achieve the aim, a large data bank, including 496 experimental samples, was compiled from published articles. Temperature (273.15 to 413.15 K), partial pressure of CO ₂ (0.0026 to 6894.76 kPa), and concentration of amine (0.5 to 8 mol/L) were utilized as input features for training models. The results demonstrated that the GEP technique, having a determination coefficient (R ² ) of 0.9506 as well as a root mean square error (RMSE) of 0.0831, outperforms other techniques in terms of the CO ₂ loading capacity prediction of DEA solutions. The R ² estimates of 0.9415 and 0.9301 for the GP as well as GMDH, accordingly, proved that these two models also produced accurate findings. In addition, it was found that the developed GEP model can accurately follow the patterns of CO ₂ loading capacity in response to changes in input variables. The finding of the last research part, the Leverage technique, revealed that the data used for modeling are valid, and the GEP approach is statistically robust. The primary contribution of this research is how interpretable machine learning techniques are combined with chemical process modeling to provide a framework for enhancing CO ₂ absorption system.

The efficient and cost-effective purification of natural gas, particularly through adsorption-based processes, is critical for energy and environmental applications. This study investigates the nitrogen (N₂) adsorption capacity across various Metal-Organic Frameworks (MOFs) using a comprehensive dataset comprising 3246 experimental measurements. To model and predict N₂ uptake behavior, four advanced machine learning algorithms—Categorical Boosting (CatBoost), Extreme Gradient Boosting (XGBoost), Deep Neural Network (DNN), and Gaussian Process Regression with Rational Quadratic Kernel (GPR-RQ)—were developed and evaluated. These models incorporate key physicochemical parameters, including temperature, pressure, pore volume, and surface area. Among the developed models, XGBoost demonstrated superior predictive accuracy, achieving the lowest root mean square error (RMSE = 0.6085), the highest coefficient of determination (R² = 0.9984), and the smallest standard deviation (SD = 0.60). Model performance was rigorously validated using statistical metrics and graphical analysis. Trend consistency with experimental data confirmed that XGBoost accurately captures the effect of pressure on N₂ uptake. Additionally, SHAP (Shapley Additive Explanations) analysis identified temperature as the most influential factor in adsorption prediction. Finally, an outlier assessment using the Leverage method indicated that approximately 94% of the data points were statistically valid and within the model’s applicability domain.

Energy and environmental policy agencies have been looking for suitable adsorbent materials to promote the use of adsorbed natural gas (ANG). Various candidate adsorbent materials have been developed and tested for methane adsorption. Metal-Organic Frameworks (MOFs) have shown a promising performance in methane adsorption and are of particular interest due to their power in adsorption and separation of gases, chemical tunability, ease of synthesis, and high surface area. Accurate calculation of the theoretical adsorption potential of methane in MOFs and its validation through experiments brings about significant challenges. A growing number of researchers are adopting soft-computing approaches, particularly machine-learning (ML) algorithms, to tackle these challenges. Although ML algorithms have been applied in assessing methane uptake capacity of MOFs, the majority of these efforts have primarily focused on feature selection or the criteria for MOF screening. This communication, however, mainly focuses on the implementation of ensemble-based ML paradigms, including gradient boosting (GBoost), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), and gradient boosting with categorical features support (CatBoost) in accurate estimation of methane uptake capacity of experimentally synthesized MOFs based on some readily available features including temperature, pressure, and MOF’s pore volume and surface area, for the first time. To this end, a database containing almost 2600 datapoints was attained. The results indicated the high performance of the XGBoost algorithm in estimating the methane uptake capacity of MOFs with a correlation coefficient (R²) of 0.9955. Moreover, further analyses revealed that the developed predictive model can reliably estimate the physical trend of CH₄ capacity variations with changing pressure. Also, further analysis indicated the large impact of pressure value on the predicted values. The employed outlier detection technique showed that almost 95% of the collected data points were valid.

Oxygen and nitrogen are essential gases with extensive applications in industrial and medical fields, making their separation from air a critical yet complex challenge. Zeolites, renowned for their highly porous structures, have garnered significant attention as promising materials for enhancing gas storage and separation processes. In this paper, advanced machine learning (ML) methods, including Generalized Regression Neural Network (GRNN), Cascade Forward Neural Network, and Multilayer Perceptron were utilized to forecast the O₂ uptake capacity in zeolites. A comprehensive database of 750 experimental O₂ uptake values was constructed, incorporating pressure, pore volume, temperature, and surface area as input features to develop robust predictive models. The findings demonstrated the superior performance of the GRNN model, achieving an exceptional root mean square error of 0.03 and a coefficient of determination (R²) of 0.9991, outperforming the other techniques. Additionally, further analyses confirmed the reliability of the presented models in accurately capturing physical trends of O₂ uptake under varying pressure conditions at different constant temperatures. Sensitivity analysis further revealed that pressure positively influences O₂ storage, while temperature exerts the most significant effect, with relevancy factors of 0.341 and - 0.848, respectively. These results underscore the effectiveness of ML techniques in precisely forecasting and enhancing gas storage processes in zeolites, offering meaningful insights to drive advancements in separation technologies.

In-situ combustion (ISC) offers a compelling solution for enhancing oil recovery, particularly for heavy crude oils. This process involves the oxidation and pyrolysis of hydrocarbons, generating heat and depositing fuel in the combustion front. In this work, the thermo-oxidative profiles and residue formation of crude oils during thermogravimetric analysis (TGA) were modeled using 3075 experimental data points from 18 crude oils with API gravities ranging from 5 to 42. Four advanced tree-based machine learning algorithms comprising gradient boosting with categorical features support (CatBoost), light gradient boosting machine (LightGBM), random forest (RF), and extreme gradient boosting (XGBoost) were utilized to develop accurate predictive models. The results indicated that CatBoost outperformed the other models, achieving mean absolute relative error (MARE) values of 4.95% for the entire dataset, 5.92% for the testing subset, and 4.71% for the training subset. Moreover, it recorded a determination coefficient (R²) value of 0.9993, highlighting its exceptional predictive capability. Furthermore, temperature was identified as the most influential factor affecting residual crude oil content, exhibiting a significant negative correlation, while API gravity also showed a negative impact. Conversely, asphaltene, resin, and heating rate positively correlated with residual content. Finally, the leverage method demonstrated that only 2.14% of the data were identified as suspected, with no out-of-leverage points detected, underscoring the reliability of the CatBoost model and the gathered experimental data. Effective management of fuel consumption and residue formation is crucial for maintaining the ISC process, and the CatBoost model has demonstrated strong predictive capabilities that support this objective.

This study focuses on predicting the heat capacity of pure liquid-phase ionic liquids (ILs) using machine learning models from various categories, including support vector machines, instance-based learning, ensemble learning, and neural networks, with linear regression serving as a baseline. A key aim of this work is not only to achieve accurate predictions but also to ensure the interpretability of the results, addressing a gap often overlooked in predictive modeling studies. To accomplish this, we curated and cleaned a comprehensive data set of 13,893 data points covering 322 ILs, using temperature and chemical structure-based features as inputs. We evaluated model performance and conducted a thorough interpretability analysis to reveal the patterns of the top-performing model's predictions, ensuring that they are understandable. All models outperformed the baseline, with XGBoost (eXtreme Gradient Boosting) from the ensemble learning category achieving the best results, with total RMSE, R², and AARD (%) values of 11.389, 0.997, and 1.212%, respectively. Shallow neural networks also performed competitively, suggesting that complex deep learning architectures may not be necessary. Both 10-fold and leave-one-IL-out (LOILO) cross-validation further validated the robustness of these results. Importantly, the interpretability analysis identified key factors influencing heat capacity predictions, such as anion size (e.g., NTf₂ and FAP) and alkyl chain length. These factors were validated by testing the model on previously unseen IL examples. Additionally, a user-friendly web application was developed to make predictions, allowing users to input chemical groups or select compounds from a predefined list of 1633 ILs. This study underscores the importance of combining diverse modeling approaches with robust interpretability techniques to achieve reliable and explainable predictions for IL heat capacity.