Development of a Predictive Model to Correlate the Chemical Structure of Amines with Their Oxidative Degradation Rate in a Post-Combustion Amine-Based CO₂ Capture Process Using Multiple Linear Regression and Machine Learning Regression Approaches

Abstract

In this study, the degradation behavior of 30 different amines was investigated, which were categorized into four distinct groups: alkanolamines, sterically hindered alkanolamines, multialkylamines, and cyclic amines. These experiments were conducted over a period of 14 days at a temperature of 60 °C, with a feed gas comprising 99.9% O₂ flowing at a rate of 200 mL/min. The primary objective was to establish a correlation between the chemical structures of these amines and their susceptibilities to degradation. To assess this, the concentration of the amines at various time points was measured to determine their degradation rates. Results showed that secondary amines exhibited degradation rates higher than those of primary and tertiary amines. Amines with cyclic structures demonstrated lower oxidative degradation rates. Longer alkyl chain lengths decreased degradation rates in all amine types because of their electronic and steric hindrance properties. A higher number of hydroxyl groups increased the degradation rate by destabilizing the free radical. An increase in hydroxyl groups in nonsterically hindered amines increased the degradation rate by decreasing free radical stability. In contrast, for sterically hindered amines, an increase in hydroxyl groups decreased the degradation rate because the steric hindrance effect is now more dominant than the electron-withdrawing effect. An increase in the number of amino groups led to higher degradation rates due to the presence of more reactive sites for free radical formation. Amines with tert-alkyl groups exhibited higher degradation rates than those with straight chains. Moreover, branched alkyl groups located between amino and hydroxyl groups significantly increased the degradation rates. Two degradation models, a semiempirical statistical model and a CatBoost machine learning regression model, were developed to predict amine degradation rates based on their chemical structure and relevant properties. To train these models, a data set of 27 different amines was used, while another set of 3 amines was reserved for testing the model’s predictive performance. The average absolute deviations (AAD) achieved were, respectively, 22.2 and 0.3%.

1. Introduction

Postcombustion CO₂ capture using an aqueous amine solution is commercial conventional technology utilized for minimizing the release of CO₂ from the power generation plant. It is known as a promising CO₂ capture technology with high maturity and reliability and great suitability and feasibility with an existing power plant, low investment, high reactivity and efficiency, reusability of solvent,¹⁻⁶ and practical use even at low partial pressure of CO₂ in the flue gas.⁷ Operational problems inevitably occur in this process. They typically occurr during the plant start-up and operation. Amine degradation is one of the serious operational problems that occur through side reactions with CO₂ and other flue gas components.⁸⁻¹⁰ It causes the reduction of process efficiency due to solvent loss that requires a fresh solvent to be made up in order to maintain the CO₂ capture performance efficiency.^11,12 In addition, the generated degradation products also cause other severe operational problems such as foaming,¹³ fouling,^14,15 increased viscosity,⁸ corrosion, and environmental issues due to waste disposal and generated volatile compounds.¹⁶⁻¹⁸ The additional cost from these problems is estimated to be about 10% of the total CO₂ capture cost.¹⁹

Generally, the degradation of any amine solution used in the CO₂ capture process occurs via thermal and oxidative degradation pathways. The first one usually occurs at high temperatures in the stripper (100–150 °C) and in the presence of CO₂.²⁰⁻²⁴ The second one mainly takes place in the presence of oxygen (O₂), NO_x, SO_x, and free metal ions under absorber conditions.¹¹ It can also occur in the rich solvent at the cross-exchanger outlet at 100–145 °C.²⁵ In the case of the oxidative degradation pathway, two main mechanisms, namely, electron abstraction and hydrogen abstraction, have been proposed in the literature.²⁶⁻³² In electron abstraction, an electron from the nitrogen atom is first abstracted by a reactive free radical to produce an ammonium radical, which further loses its free radical to produce an imine radical. The hydrolysis of the imine radical then further occurs to produce an aldehyde.^27,33 In hydrogen abstraction, the hydrogen is likely extracted from an α- and/or β-carbon to the nitrogen atom and also from the nitrogen itself. Then, an internal and external amine radical transfer is performed to produce ammonium aldehydes and aldehyde radicals as the main degradation products.^27,24−35Figure 1 shows a simplified initial step of the radical-based oxidative degradation pathway of the amine.

Figure 1.

Widely accepted mechanism step that triggers amine degradation.

The intensity of oxidative degradation not only depends on gas composition but also depends on the structure of the amine solvent itself. The extent of the oxidative degradation mechanism is primarily influenced by the electronic and steric hindrance properties inherent in the amine’s molecular structure. When an electron-withdrawing functional group surrounds the free radical center, it exhibits the capacity to draw electrons away from both the N and the α-C radical centers. This electron withdrawal destabilizes the free radical and weakens the N–C bond, consequently rendering the amine more susceptible to degradation. On the other hand, if the amine contains groups with electron donation ability that provide electrons, it will strengthen the N–C bond, thereby reducing the chances of the amine generating a free radical needed for further degradation.³⁶ In the case of steric hindrance, varying degrees of degradation protection can be obtained depending on how steric the chemical structures are. A more steric structure provides greater stability against the formation of amine free radicals than a less steric substituent group, thus making the amine more difficult to degrade.³⁶ Based on this mechanism, it implies that an amine with a structural feature connected to the amino group with the chemical property that favors the formation of an amine radical will likely cause degradation to occur more easily than amines with a substituent containing a structure that does not chemically favor the formation of the radical. This is because the chemical property and reactivity of the amine’s substituent are a manifestation of its own atomic and molecular connectivity. This further implies that the degradation potential of the amine is highly dependent on the type of substituent surrounding the amine reaction center.

This dependency has been confirmed by various researchers who have shown that different amine types exhibit different vulnerabilities for oxidative degradation. In 2009, Lepaumier et al.³⁷ studied the degradation of 12 different amines at 140 °C using 4 M of amine solution in a stainless steel batch reactor for 15 days using 0.42 MPa of O₂ pressure. They found that secondary amines were less stable than primary and tertiary amines due to their highest nucleophilicity to form the addition products (dimers) by intermolecular substitution of alcohol with amine. This dimer can form ring closure by intramolecular nucleophilic substitution. Tertiary amines require preliminary demethylation or dealkylation to initiate the degradation reaction, and the stability of hindered primary amines is generally close to the tertiary amines, for which the steric hindrance structure prevents the oxazolidone ring opening mechanism into addition products. In 2010, this research group also observed that some tertiary polyamines [N,N,N’,N′- tetramethylbutylenediamine (TMBDA) and N,N,N′,N′,N″-pentamethyldiethylenetriamine (PMDETA)] are much less stable than MEA due to the ability of the alkyl chain length between amino groups to easily produce stable five- or six-membered ring formation.³⁸ In 2013, Vevelstad and co-workers also studied the oxidative degradation of five amines in a closed batch system. They found that the order of increasing degradation rate was dimethylamino ethanol (DMAE) (tertiary) < MEA (primary) < 2-(methylamino) ethanol (MAE) (secondary).³⁹ Furthermore, the research conducted by Buvik and colleagues⁴⁰ investigated the importance of steric effects, the presence of CO₂, and various steric factors when assessing the stability of amines under oxidative conditions. The relationship between biodegradability and oxidative degradation yielded valuable insights into the degradation patterns of amines.

As shown in the literature,³⁷⁻⁴⁰ the degradation of amines hinges on the arrangement of their functional groups within the amine structure. Consequently, understanding the relationship between the chemical structure and amine degradation is crucial to an accurate estimation of the degradation rate of any unknown amine. This prior knowledge of amine degradation not only aids in the selection of amines that are less prone to degradation during design but also enables the development of an effective plan to prevent degradation, thereby minimizing its impact in an amine-based CO₂ capture process. In 2012, Martin and colleagues⁴¹ conducted a study focused on predicting amine degradation by constructing a quantitative structure–property relationship (QSPR) model. The degradation of 22 amine solutions was assessed in stainless-steel batch reactors at 140 °C for 14 days under a pressure of 0.5 MPa, with a mixture of 75% CO₂, 20% N₂, and 5% O₂. The developed QSPR model incorporated a straightforward molecular descriptor that included calculated dipole moment, calculated pK_a value, and topology and Jurs descriptors, all within Material Studio. The predictive model, based on data from 15 different amines, demonstrated a high level of accuracy with an R² value of 0.85 at a 95% confidence level when comparing experimental and predicted data values. This model proved to be valuable in assisting the selection of efficient solvents for CO₂ capture. In 2020, Raznahan et al.⁴² took a similar QSPR approach, developing both linear and nonlinear models to predict the absorption capacity and absorption rate of CO₂ in amine solutions. These models were constructed using 11 non-ring-shaped amines and 17 ring-shaped amines. The nonlinear model outperformed the linear one, exhibiting higher efficiency and accuracy with R² values of 0.895 and 0.954 for non-ring- and ring-shaped amines, respectively, and a %error between calculated and experimental values of less than 10%. Moreover, this model was employed for the design of new amine compounds with elevated absorption capacity and absorption rate, further enhancing its utility in the field.

The present study investigated the relationship between the chemical structure of amines involving the presence, type, arrangement, and positioning of functional groups such as amino, alkyl, and hydroxyl groups in both noncyclic and cyclic amines with their degradation rates. The information obtained was employed to develop a predictive degradation rate model using a semiempirical multiple linear regression and further improved the model with a Categorical Boosting (CatBoost) machine learning regression approach. CatBoost regression is one of the deep learning methods used in regression tasks. The method adjusts the weights of each tree using gradient boosting and categorical features, which allows the model to handle categorical data well without having to convert it to numerical values, also known as encoding.⁴³ A key advantage of CatBoost regression is its ability to handle overfitting effectively. This is achieved through the use of regularization to reduce the complexity of trees and early stopping techniques to halt the training of the model when unnecessary, thus increasing the model’s predictive performance. In updating the weights of CatBoost, a technique called gradient-based one-sided sampling (GOSS) is used. This involves calculating the gradient of the samples and using it to rank them. Higher weights are assigned to samples with larger gradients and lower weights to samples with smaller gradients. These weights are then used in the loss calculation process before updating the weights of the various variables used in the model.⁴⁴ Incorporating CatBoost into our analysis enhances the predictive capabilities of our model, complementing the results obtained from multiple linear regression. This hybrid approach allows us to benefit from the strengths of both methods while still preserving the valuable insights gained from the linear regression analysis. The combination of CatBoost and multiple linear regression provides a more comprehensive understanding of the underlying relationships between the features and target variable, leading to more accurate predictions of amine degradation rates. The unique aspect of this study is the use of a hyphenated approach, whereby the machine learning approach is used to achieve improved predictability, following a conventional regression methodology which is used to relate the degradation rate with an amine structure with reasonable predictive accuracy. This novel approach is specifically useful in selecting the least degradable solvent to use for the design of an amine-based CO₂ capture plant without resorting to laboratory experiments. This model provides a convenient tool for evaluating the degradation potential of selected amines or amine blends intended for use in an amine-based CO₂ capture plant. The predicted degradation rates obtained from the model equations can be employed effectively in the advanced planning and design of a solvent degradation management program, aiding in the development of an efficient and sustainable amine-based CO₂ capture process.

2. Experimental Section

2.1. Chemicals and Equipment

Thirty amines in four categories were used. The names, categories, and chemical structures are shown in Figure 2. They were obtained mostly from Sigma-Aldrich. Each amine solution was prepared to the concentration of 2 mol/L (M), and the actual concentration was confirmed by titration with 1 N HCl (Fisher Scientific) using methyl orange (Sigma-Aldrich) as an indicator. Industrial grade O₂ (purity of 99.5%) used for the oxidative degradation test was purchased from Praxair Canada. Minitab software (Minitab 18 by Minitab LLC, USA) was used to estimate the coefficients for the predictive model of oxidative degradation.

Figure 2.

Chemical structures of (a) nonsterically hindered alkanolamines, (b) sterically hindered alkanolamines, (c) multialkylamines, and (d) cyclic-amines.

2.2. Amine Degradation Test

This study is primarily concerned with elucidating the intrinsic impacts of various amine structural configurations on the oxidative degradation of amines. Therefore, O₂ was singularly employed as the primary flue gas component for the degradation experiment. This experimental design was specifically chosen to facilitate a comprehensive evaluation of the genuine influence of amine structural diversity on oxidative degradation, uncontaminated by extra factors such as CO₂, SO_x, and NO_x, and metal constituents found in fly ash. In addition, 99.5% of O₂ was employed to accelerate the reaction, allowing for the observation of degradation trends within a 14-day experimental time frame. The results derived from these experiments subsequently serve as the basis for substantive relationships between oxidative degradation and the chemical structure of the amines under investigation. To maintain experimental uniformity and minimize potential issues arising from precipitation and high viscosity that are characteristic of certain aqueous amine solutions, a concentration of 2 M was selected for all of the examined amines. According to our evaluation, although the actual values of degradation rates may vary at concentrations other than 2 M, the comparative trends in degradation rates would closely resemble those documented in this study. The schematic diagram of the experimental setup for oxidative degradation of amine solution is shown in Figure 3. Exactly 250 mL of the 2 M amine solution was placed in the reaction flask and warmed to 60 ± 2 °C before the experiment was started. O₂ of 99.5% purity at a flow rate of 200 ± 2 mL/min, manually regulated with a needle valve and measured by a rotameter (Aalborg, model P single flow tube, 0–150 mm with a range of 0–500 mL/min ±1.0% error) and bubble flow meter (Agilent Optiflow 420 Model), to ensure that the flow rate was constant throughout the experiment, was used as the feed gas. This was bubbled through a water saturator prior to its introduction into the reaction flask for the oxidative degradation reaction. The temperature of the outlet stream from the condenser was made to be equal to the temperature of the inlet feed gas using a circulating cooling water bath in order to ensure water balance. The degradation reaction started once the feed gas was bubbled into the amine solution. The reaction continued for 14 days, with an amine solution sample taken once a day. The collected amine solution samples were placed in the refrigerator before GC-MS analysis to minimize further amine degradation. All the amine samples taken were analyzed to determine the concentration of amine left in the reaction chamber using the GC-MS technique. Details of this analysis are given in our previous study.⁴⁵

Figure 3.

Experimental setup for the amine degradation test.

The concentration of the amine solution was quantified through a calibration curve, which was established by using various concentrations of standard amine solutions, prepared individually for each amine type. The preparation of these standard samples for each amine was conducted meticulously and with precision, ensuring coverage of the concentration range from 0.05 to 0.5 M. Furthermore, an internal standard of 0.2 M ethylene glycol was introduced to mitigate potential variations in GC-MS conditions and run-to-run analyses, thereby enhancing the reliability and accuracy of the results. To ensure robustness and accuracy, both the standard samples and all degraded samples underwent a three-time repetition of analysis, followed by the calculation of the standard deviation. Periodic verification of the calibration curve using known concentrations of amine solutions was also carried out to ensure its accuracy, a critical factor that, in turn, determined the accuracy of the degraded amine concentrations. Consequently, the aggregate error for the degradation analysis was ultimately calculated to be 3.5%. The degradation rate for each amine solution was ascertained by determining the slope of the concentration–time plot. This analytical approach not only enabled us to quantify the degradation rates of distinct amines but also facilitated comparative evaluations, thereby yielding valuable insights into their degradation behaviors.

2.3. Predictive Model Development Using Multilinear Regression and CatBoost Machine Learning

Experimental degradation rate data for all 27 selected amines were employed as a foundational data set to thoroughly construct a predictive model. This model has been meticulously designed to extrapolate the intricate relationships between amine molecular structures and their corresponding degradation rates, thus offering a convenient tool for comprehensively assessing the stability of these amines in the context of amine-based CO₂ capture processes.

2.3.1. Identifying Variables for Predictive Model Development

The amine degradation rate model was formulated based on its chemical structure. The model was made up of the amine’s structural and reaction variables. The structural variable considered different chemical components commonly found in amines, which are used for carbon capture and used to represent numbers of all the possible groups present within the amine structure. The structural variables consisted of (1) all amino groups in the noncyclic structure: primary amino group (−NH₂−), secondary amino group (−NH<), tertiary amino group (>N< ), (2) all amino groups in the cyclic structure: secondary amino group (−NHc<), tertiary amino group (>Nc<), (3) all alkyl groups in the noncyclic structure: primary carbon (−CH₃), secondary carbon (−CH₂−), tertiary carbon (−CH<), and quaternary carbon (>C<), and (4) secondary carbon in the cyclic structure (−CH₂c−), 5) hydroxyl group (−OH−). The reaction variables took into account the chemical connectivity between different structural groups present within the amine molecule that could either enhance or slow the formation of the amine radical. These variables were (1) electron-withdrawing parameter (EWG), (2) electron-donating parameter (EDG), (3) steric hindrance parameter (S), and (4) interaction parameter (I).

Figure 4a shows the effect of an EDG substituent in which R₁ has the ability to donate its electrons to the central N atom, increasing the electron density in the vicinity of its carbons close to the N, which in turn decreases the probability of the amine in forming the radical, a step required to trigger degradation. Even if the radical is formed, its stability is enhanced by the aforementioned phenomenon; hence, the degradation reaction is even harder to occur.³⁶ An EWG group, however, has the opposite effect on radical formation. As shown in Figure 4b, the EWG substituent (also represented by R₁) has the ability to attract the bonding electrons away from the central N atom, thus increasing the richness of electron density over the carbons close to the N. This increases the amine’s radical formation probability, thereby promoting degradation. The radical can also be destabilized further by the presence of the EWG group, further causing the amine to degrade even more easily.³⁶ This is then taken as the effect that opposes the degradation of the amine.

Figure 4.

Electronic effects of amine-substituting groups on amine degradation: (a) electron-donating group effect and (b) electron-withdrawing group effect.

The steric hindrance parameter (S) can be explained using Figure 5a,b, which compare two amines with less steric and more steric substituting groups, respectively. Both groups oppose the degradation of the amine, as they both minimize the amine’s radical formation. The interaction parameter (I) was introduced into the degradation potential equation in order to take into account a possible synergy between all the reaction parameters (i.e., EWG, EDG, and S parameters).

Figure 5.

Steric hindrance effect of amine-substituting groups on amine degradation: Less steric group and (b) more steric group.

2.3.2. Determination of the Degradation Parameter Values

Structural variable values were simply determined by counting the number of substituents present in the chemical structure of a given amine. For example, MEA consists of 1 −NH₂–, 0 −NH<, 0 >N< , 0 −NHc<, 0 >Nc<, 0 CH₃, 2 −CH₂–, 0 −CH<, 0 >C<, 0 −CH₂c–, and 1 −OH–. EWG and EDG variable values were calculated with the electronegativity value (EN) of N taken as a reaction center of the molecule, which is either subtracted or added, respectively, by the sum of the electronegativity of all the substituents surrounding the nitrogen reaction center. EWG and EDG values for each amine category were calculated using the following equations (eqs 1–6):

1

2

3

4

5

6

where n_N, n_C, n_H, and n_O represent the count of nitrogen (N), carbon (C), hydrogen (H), and oxygen (O) atoms within the amine structures. The electronegativity (EN) values for N, C, H, and O are 3, 2.5, 2.1, and 3.5, respectively.⁴⁶ Moreover, in the case of cyclic amines, the nitrogen atom to which the side chain is attached is referred to as “N-para” within this study. Examples of calculations for EDG and EWG values can be found in the supporting document.

Steric hindrance (S) values of all amines were calculated by combining A values of all of the chemical groups present in the amines. The A value assigned to each substituent in an amine structure was taken directly from the literature,^47,48 which represented the thermodynamic value (e.g., Gibbs free energy) that was used to show the degree of the spontaneity of the reaction of the amine containing such a substituent. This could, in turn, also represent steric hindrance of the substituents on the amine structure. The chemical groups used in this work were methyl, ethyl, propyl, n-butyl, t-butyl, pentyl, hexyl, hydroxyl, primary amino, secondary amino, and tertiary amino groups with the corresponding A values of 1.7, 1.75, 1.8, 1.85, 4, 1.9, 1.95, 0.87, 1.6, 1, and 2.1, respectively. It must be noted that the A values of propyl, n-butyl, pentyl, and hexyl were estimated from those of methyl and ethyl. S values of all amines were calculated by using eq 7.

7

where n_amino, n_alkyl, and n_hydroxyl denote the number of amino, alkyl, and hydroxyl functional groups in the amine structure, while A_amino, A_alkyl, and A_hydroxyl represent the corresponding A values for those specific chemical groups in the given amine structure. Examples of calculations for S values can be found in the supporting document.

In statistical analysis, an interaction variable constitutes a distinctive feature that encompasses three or more variables. This characteristic manifests itself when the interplay of two or more variables exerts an impact on a third variable in a nonadditive manner. The salience of interaction variables in statistical models resides in their ability to accommodate intricate relationships and comprehend the nuances of how diverse variables mutually influence one another. Including interaction terms augments the flexibility inherent in specifying a linear model, permitting the delineation of distinct slopes for different lines. This increased flexibility contributes to a more precise representation of the data and enhances predictive performance.⁴⁹ Consequently, in the context of this study, the interaction variable was generated based on the reaction variable directly affecting the oxidative degradation reaction involving EDG, EWG, and S. Specifically, EWG accelerates the degradation rate while EDG and S diminish the degradation rate. As used in our formulation, variables that increase the degradation rate constitute the numerator, while those decreasing it form the denominator. The value of the interaction variable is derived from the combination of these variables, as illustrated in eq 8.

The interaction parameter was calculated using eq 8.

8

All of the model parameters of the other amines were calculated similarly to those of MEA. The model parameter values are tabulated in Table 1.

Table 1. Degradation Model Parameter Values.

amine	structural variables											reaction variables			interaction variable, I
	NH2	NH	N	NHc	Nc	CH3	CH2	CH	C	CH2c	OH	EWG	EDG	S
MEA	1	0	0	0	0	0	2	0	0	0	1	–9.84	13.36	4.22	–0.17
DEA	0	1	0	0	0	0	4	0	0	0	2	–26.96	13.14	6.14	–0.33
EAE	0	1	0	0	0	1	3	0	0	0	1	–9.84	26.76	5.37	–0.07
BAE	0	1	0	0	0	1	5	0	0	0	1	–9.84	40.16	5.47	–0.04
MDEA	0	0	1	0	0	1	4	0	0	0	2	–26.96	19.84	9.04	–0.15
3A1P	1	0	0	0	0	0	3	0	0	0	1	–13.45	16.45	4.27	–0.19
3DMA1P	0	0	1	0	0	2	3	0	0	0	1	–13.45	29.85	8.17	–0.06
AHMPD	1	0	0	0	0	0	3	0	1	0	3	–31.43	12.17	8.21	–0.31
t-BEA	0	1	0	0	0	3	2	0	1	0	1	–9.84	40.16	7.62	–0.03
1DMA2P	0	0	1	0	0	3	1	1	0	0	1	–13.45	29.85	9.82	–0.05
PZ	0	0	0	2	0	0	0	0	0	4	0	6.00	23.60	5.50	0.05
EPZ	0	0	0	1	1	1	1	0	0	4	0	6.00	37.00	8.35	0.02
MPZ	0	0	0	1	1	1	0	0	0	4	0	6.00	30.30	8.30	0.02
EAPZ	1	0	0	1	1	0	2	0	0	4	0	6.00	35.40	9.95	0.02
5A1P	1	0	0	0	0	0	5	0	0	0	1	–20.69	22.61	4.37	–0.21
t-BDEA	0	0	1	0	0	3	4	0	1	0	2	–26.96	39.94	11.34	–0.06
DETA	2	1	0	0	0	0	4	0	0	0	0	9.00	32.90	7.70	0.04
DAP	2	0	0	0	0	0	3	0	0	0	0	6.00	24.45	5.00	0.05
HMDA	2	0	0	0	0	0	6	0	0	0	0	6.00	34.50	5.15	0.03
TEA	0	0	1	0	0	0	6	0	0	0	3	–45.16	11.84	9.96	–0.38
DMAE	0	0	1	0	0	2	2	0	0	0	1	–9.84	26.76	8.12	–0.05
BDEA	0	0	1	0	0	1	7	0	0	0	2	–26.96	39.94	9.19	–0.07
EDA	2	0	0	0	0	0	2	0	0	0	0	6.00	21.10	4.95	0.06
Proprietary amine	0	0	0	1	1	0	2	0	0	4	1	–6.84	27.66	9.67	–0.03
4DMA1B	0	0	1	0	0	2	4	0	0	0	1	–17.07	32.93	8.22	–0.06
EEDA	1	1	0	0	0	1	3	0	0	0	0	6.00	34.50	6.10	0.03
EDEA	0	0	1	0	0	1	5	0	0	0	2	–26.96	26.54	9.09	–0.11

This work has also addressed the issue of vast differences in the values of the variables, which could affect the model’s predictive ability of the degradation rate and, consequently, the prediction accuracy. First, a factor of 10 was used to divide the values of EWG, EDG, and S variables, which were considered to have large value differences. This step was carried out prior to regression analysis to normalize the data points so that they fell within the order of magnitude close to those of the structural and interaction variables. The normalized data set is shown in Table S1. The normalized data were then subjected to a multilinear regression analysis similar to what was used on the original data set provided in Table 1. The coefficients from the regression for all the 15 variables in eq 9 were then compared to those obtained from the original data set, as shown in Table S2. Based on our assessment, the large differences in the values did not affect the regression outcomes. This was evident from all the variables, in that the corresponding variables had exactly the same coefficients and standard errors, with minor exceptions being seen only in EWG and interaction variables (i.e., 2 out of 15 variables). In addition, the two coefficient sets both predicted exactly the same rate values for all the amines used in generating the model, and thus the model predictability was not affected. Based on these, it was decided to proceed with the original data set without altering it so that the final degradation rate model could be obtained for this study.

2.3.3. Degradation Model Development Using Multilinear Regression

The power law serves as a common choice for elucidating the kinetics of chemical reactions.⁵⁰ Consequently, the power law equation was deemed to be a suitable initial framework for formulating the degradation rate models in this study. The power law equation was first used to correlate all parameters to the degradation rate measured in terms of the initial degradation rate of amine (mM/h). The overall degradation rate model for the prediction of amine degradation rate as given in eq 9 is thus:

9

where the degradation rate represents the initial rate of amine degradation (mol/L h). The a to o are the exponents of the structural, reaction, and interaction variables that generally indicate the sensitivities in changing their values to the change of the amine degradation rate. Multiple linear regression was used to determine the degradation model parameter exponents of a to o and the coefficient of K, which are shown in eq 9. Data given in Table 1 were used for regression. However, the structural variable data specific to those of no value (values being zero due to the absence of certain chemical groups with the amine structure) were inputted as 1 × 10^–10 to run the regression successfully to provide numerical values of the unknowns that are useful for the degradation model equation.

2.3.4. Degradation Model Development Using CatBoost Machine Learning

In this work, CatBoost was further used to determine the relationship between the degradation rates and all 15 independent variables shown in Table 1. This refined approach integrated the machine learning technique to achieve a more comprehensive amine degradation predictive model. Following extensive experimentation with various methods, it was determined that the results obtained were unsatisfactory due to the limited amount of original data available for analysis. In order to address this issue, the Synthetic Minority Oversampling Technique was employed to generate additional data points by interpolating between existing ones based on their proximity. A total of 24 synthetic data points were obtained through this method. Subsequently, the original data set was combined with this augmented data. Notable significant improvements in the overall outcomes were achieved by utilizing CatBoost which facilitated gradient boosting on decision trees.

These methodologies were iteratively applied to the construction of ensemble models. In the initial iteration, the algorithm obtained the primary tree with the intent to minimize the training error. However, it is worth noting that this model often displayed significant errors; building excessively large trees in boosting is discouraged due to the risk of overfitting the data. Moving to the second iteration, the algorithm learned an additional tree to alleviate the error introduced by the first tree. This process was consistently repeated until the algorithm generated a model of satisfactory quality.⁵¹

3. Results and Discussion

3.1. Amine Structural Correlations on Amine Degradation

The studied amines in each category were chosen for fair comparison by first identifying the benchmark amine within each group. Additional amines were then included alongside the benchmark amine in each category to ensure a diverse range of structural variations for evaluation. For instance, in the primary alkanolamine series, MEA was initially selected, as it is a commonly used alkanolamine for CO₂ absorption in commercial applications. This series was further expanded to encompass other amines, including those with longer alkyl chain lengths between functional groups (such as 3-amino-1-propanol), differing substitutions of alkyl groups at the nitrogen atom (N-substitution; like MAE), alternative functionalities aside from hydroxyl groups (e.g., diethanolamine), and compounds with varying steric hindrance (such as AMP). A similar selection process was employed for multialkylamines. In the case of cyclic amines, derivatives of piperazine (PZ) were chosen due to their widespread use in CO₂ capture studies, which encompass investigations into absorption efficiency as well as thermal and chemical stability (as cited in refs (52−55)).

3.1.1. Effect of Being Primary, Secondary, and Tertiary Amines

Figure 6 shows the effect of amine types as primary, secondary, and tertiary amino groups on the amine degradation rate. In the primary amine group represented by MEA, it showed a degradation rate of 0.99 mM/h. When one methyl (−CH₃) group was attached to the nitrogen atom as in MAE and became a secondary amine, the degradation rate increased to 2.07 mM/h. Attaching an alkyl group to the nitrogen atom, which changes the amine type from primary to secondary, increases the nucleophilicity of the structure, resulting in greater reactivity toward other additional reactions of amines with the obtained degradation products derived from the oxidative degradation reaction.³¹ When one more −CH₃ was added to the nitrogen atom as in DMAE to become a tertiary amine, the degradation rate decreased to 1.33 mM/h, which is lower than MAE but still higher than MEA. This observation was also noted in the study by Lepaumier et al.,³¹ where they reported the nucleophilicity order as follows: secondary amines > primary amines > tertiary amines. Based on our finding, the order of reactivity toward oxidative reaction was also secondary > tertiary > primary amines in the case of those amine types obtained by methyl group substitution at the nitrogen atom. The contradicted order of primary and tertiary amines was also observed in the group of 3A1P (primary) and 3DMA1P (tertiary), which showed degradation rates of 0.76 and 1.06 mM/h, respectively. This might be explained by the more favorable phenomenon of demethylation reaction that occurred in DMAE and 3DMA1P than that of dealkylamine that occurred in MEA and 3A1P. Moreover, nonsymmetrical amines, exemplified by R–N–R’ (example: R = CH₃ and R’ = CH₂CH₂OH), have the capacity to yield two distinct amino radicals, consequently leading to competitive dealkylation or radical combination processes. These contributions, which enhance the likelihood of the reaction, can further augment the superior oxidative reactivity, as discussed in prior research.⁵⁶⁻⁵⁸ This might be one of the reasons why tertiary DMAE and 3DMA1P showed a higher degradation rate than primary MEA and 3A1P in this work. Further research is still needed to arrive at a conclusive reaction mechanism.

Figure 6.

Effect of amine types as primary, secondary, and tertiary amino groups on the amine degradation rate.

When considering primary, secondary, and tertiary amines obtained through the substitution of one or two hydroxyethyl groups (−CH₂CH₂OH) on the nitrogen atom of MEA (primary), resulting in DEA (secondary) and TEA (tertiary), distinct degradation rates were observed. MEA exhibited a degradation rate of 0.99 mM/h, whereas DEA and TEA showed rates of 2.31 and 0.57 mM/h, respectively. Notably, TEA displayed the lowest degradation rate, which is attributed to its pronounced steric hindrance effect. Furthermore, when one −CH₂CH₂OH group in TEA was replaced with a methyl group, yielding MDEA (tertiary), the degradation rate increased to 0.91 mM/h. This suggests that the reactivity of amines in oxidative reactions is influenced by both electronic and steric hindrance effects.

In the case of alkyldiamines, EDA containing two primary amine groups showed a degradation rate of 1.08 mM/h. When one of them became a secondary amine by attaching a −CH₃ group to the nitrogen atom as in MEDA, the degradation rate increased significantly up to 5.53 mM/h. As described in the work of Lepaumier et al.,³⁷ the oxidative degradation of ethylenediamines can occur through dealkylation, alkylation, and addition reactions, which readily occur in such an environment. In addition, ethylenediamines with the secondary amine type are less stable and greatly favorable to forming stable imidazolidones, resulting in extremely high degradation rates.

This trend was also observed in cyclic amines. Piperazine, which contains two secondary amines, showed a degradation rate of 0.84 mM/h. The degradation rate decreased to 0.35 mM/h when a −CH₃ group was attached to one of the nitrogen atoms (as in MPZ) to generate a tertiary amine. Furthermore, as seen in Figure 3, the degradation rates of cyclic amines (piperazine and its derivative) were much lower compared with the straight chain alkanolamines and alkyldiamines. This is due to the well-known better thermal and oxidation stability of piperazine derivatives in the presence of O₂, CO₂, and metal ions.⁵²⁻⁵⁵ From the results in the literature,^37,52−54 it was seen that secondary amines are the least stable to oxidative reaction due to their higher nucleophilicity. This is followed by tertiary amines and then primary amines.^24,29

3.1.2. Effect of the Alkyl Chain Length

Figure 7 shows the trend of the degradation rate as affected by the alkyl chain length in the amine structure. Alkyl chain lengths in different positions were studied: the alkyl group located in between two nucleophilic functional groups (amino–amino and amino-hydroxyl) as well as those attached to the nitrogen atom (end alkyl). In the effect of alkyl groups between two nucleophilic groups, groups of primary alkanolamines, tertiary alkanolamines, and alkylenediamines were considered. As seen in Figure 7a, the degradation rate of primary and tertiary alkanolamines decreased with the increasing alkyl chain length in between amino and hydroxyl groups. This is because alkyl groups act as an electron-donating group, which strengthens the C–N bond, thereby reducing free radical formation and resulting in a decrease in the ability of the amine to degrade. In the case of the alkyl chain length in between two amino groups, it was observed that the degradation rate tended to decrease as the alkyl chain increased from two carbons as in EDA to three carbons in DAP. When the alkyl chain increased further to four carbons as in DAB, the degradation rate vastly increased to 1.64 mM/h and then decreased to 0.54 mM/h when the hexyl (−CH₂CH₂CH₂CH₂CH₂CH₂−) group was present as in HMDA. DAB contains the butyl (−CH₂CH₂CH₂CH₂−) group in between two amino groups, which could potentially degrade into a five-membered ring compound and ammonia by intermolecular interaction.³⁷ This observation corroborates the study of Lepaumier et al.,³⁸ which observed that TMBDA and PMDETA that were in the same family as DAB and DETA could give five- and six-membered rings, respectively, and exhibited a very high rate of degradation. However, even though HMDA also has the potential to form a seven-membered ring structure, this is less favored than five- and six-membered rings. This is because a seven-membered ring forces the bonds into a wider angle than the normal sp³ tetrahedral angle, leading to straining of the bond angle.⁵⁹ In this amine, a decrease in the degradation rate was observed mainly because of the increase in the alkyl chain length that strengthens the C–N bond by increasing free radical stability.

Figure 7.

Effect of alkyl chain length (a) between two nucleophilic groups and (b) attached at the nitrogen atom (end alkyl group) on an amine degradation rate.

An increase in alkyl chain length at the end group position (attached to the nitrogen atom) also shows a decreasing trend (see Figure 7b) in secondary alkylalkanolamines, alkylethylenediamines, and alkyl-cyclic amines. The degradation rate was observed to increase as the alkyl length on the amines became shorter. In alkyl-tertiary amine groups (MDEA, EDEA, and BDEA), it was found that the degradation rate decreased when the end alkyl chain length was increased from 1 carbon (MDEA) to 2 carbons (EDEA) but increased when the number carbons reached 4, as in BDEA. The decrease of the degradation rate as a result of alkyl chain length is not only due to the electronic effect that strengthens the N–C bond by electron donation, but also due to the enhancement of steric hindrance derived from the longer alkyl chain length that prevents free radical formation. These two factors are responsible for the amine with such a structure to degrade more slowly.

3.1.3. Effect of the Number of Hydroxyl Groups

Figure 8 shows the trend of the degradation rate as affected by the number of hydroxyl groups in three different amine types, namely, sterically hindered primary alkanolamines, secondary alkanolamines, and tertiary alkanolamines. A hydroxyl group that has the ability to reduce the free radical stability, thus weakening the N–C bond of the amine molecule and promoting degradation reactions, is considered to be an electron-withdrawing group. As seen in groups of secondary and tertiary alkanolamines, the degradation rate increased with an increase in the number of hydroxyl groups. EAE and DEA possessing 1 and 2 hydroxyl groups, respectively, exhibited degradation rates of 2.08 and 2.31 mM/h, respectively. On the other hand, EDEA and TEA, featuring 2 and 3 hydroxyl groups, respectively, displayed degradation rates of 0.29 and 0.57 mM/h, respectively. A different trend was observed for sterically hindered primary alkanolamines of AMP (0.68 mM/h) and AHMPD (0.53 mM/h), for which the degradation rate decreased with an increase in the number of hydroxyl groups from 1 to 3 groups. In this case, having more hydroxyl groups in an amine structure also makes the amine more bulky and steric, thereby making it more resistant to degradation. It is likely that in this case, the steric effect is more dominant than the electronic effect.

Figure 8.

Effect of the hydroxyl group number on the amine degradation rate.

3.1.4. Effect of the Number of Amino Groups

Amines in the groups of multialkylamines and cyclic amines were investigated to evaluate how the number of amino groups would impact degradation. Results (see Figure 9) showed an increase in the degradation rate as a result of an increase in the number of amino groups in multialkylamines. The degradation rate of EEDA was 3.03 mM/h. When one more amino group was attached to the structure to become DETA, the degradation rate drastically increased to 6.84 mM/h. The increase was because having more amino groups increased the N–C bond reaction sites for the amine to be degraded. The higher degradation rate of DETA was not only due to the more reaction sites in its structure, but also due to the tendency of the free radical in DETA to form a stable six-membered ring structure, as discussed earlier.³⁰ An increase in the degradation rate was also observed in cyclic amines, with the degradation rate of EPZ being 0.04 mM/h, which increased to 0.1 mM/h when one more primary amino group was attached to the structure as in EAPZ. However, due to the inherent chemical and thermal stability of cyclic amines,⁵²⁻⁵⁵ the increase in the degradation rate as a result of the number of amino groups they contained was not significant.

Figure 9.

Effect of the amino group number on the amine degradation rate.

3.1.5. Effect of Steric Hindrance of Alkyl Groups

The effect of the steric alkyl group was also investigated in this study by conducting a comparative analysis between straight- and branched-chain amines with identical molecular weights. Specifically, BAE and t-BAE, each with a molecular weight equal to 117.19 g/mol, were compared. Similarly, BDEA and t-BDEA, each with a molecular weight equal to 161.24 g/mol, and 3DMA1P and 1DMA2P, each with a molecular weight of 103.16 g/mol, were compared. In the case of BAE and BDEA, denoting secondary and tertiary alkanolamines, respectively, a linear butyl group is connected to the nitrogen atom. In contrast, tert-BAE and tert-BDEA feature a branched tertiary butyl group attached to the nitrogen atom. In 3DMA1P and 1DMA2P, the latter resulted from the branching out of one carbon adjacent to the OH group in 3DMA1P. As illustrated in Figure 10, the degradation rate of the straight-chain amine was lower than that of its branched-chain counterpart. Notably, within the tertiary monopropanolamines group, the degradation rate experienced a significant increase from 1.06 to 6.47 mM/h when one carbon adjacent to the OH group in 3DMA1P was moved to branch out from the main structure, forming 1DMA2P. This heightened degradation rate was attributed to the steric structure. This trend was also observed in secondary monoethanolamines and tertiary diethanolamine. Specifically, BAE and t-BAE exhibited degradation rates of 1.61 and 1.92 mM/h, respectively, while BDEA and t-BDEA showed degradation rates of 0.57 and 0.92 mM/h, respectively. Consistent with previous findings by Buvik et al.,⁴⁰ sterically hindered amines demonstrated lower oxidative stability. The degradation stability of 1DMA2P decreased due to the presence of a hydroxyl substituent at the β-position to the nitrogen atom, unfavorably influencing the amine’s stability toward degradation.

Figure 10.

Effect of the steric hindrance structure on the amine degradation rate.

3.2. Predictive Model for Amine Degradation Potential in Amine-Based CO₂ Capture Process Using Multiple Linear Regression and Catboost Machine Learning

The principal aim of this predictive model is to function as an initial tool for assessing the degradation rate of specific amines, specifically those intended for use in amine-based CO₂ capture plants. The degradation rate values obtained from this model are confined to the specific conditions delineated in this work. Nevertheless, these values prove valuable for comparative analyses across various amines and for predicting the stability of an unknown amine, in comparison to a benchmark amine. In addition, this study focuses on enhancing the accuracy of an existing semiempirical model by strategically integrating advanced machine learning techniques, aiming to improve predictive performance in the relevant field.

A multiple linear regression was performed on all 15 variables listed in Table 1. The initial regression analysis revealed that the standard errors of the −NHc<, >Nc<, CH2c–, −OH, and EDG variables were higher than their respective values. This result implied that such parameters had no significant contribution to the degradation rate prediction. Insufficient variation in the data of the aforementioned variables was probably the cause of this occurrence. As a result, a decision was made to eliminate NHc<, >Nc<, CH₂c–, −OH, and EDG variables from the model equation previously given in eq 9. This resulted in eq 10. However, the effects of the eliminated variables were still there, as shown in eq 10, as they had already been embedded in the calculations of EWG, S, and I parameters, as explained previously. A reassessment of their calculations led to the final degradation potential model given in eq 10.

10

The regression of the final model equation showed an R² of 0.92. The coefficients of the model parameters and their corresponding errors were all lower than their actual values, as given in Table 2. It is clear that the parameters used in the degradation rate equation were statistically significant and could be used to predict the degradation rate of amines only from their structures. Figure 11 shows the variance of degradation rates obtained experimentally and calculated using the models of all 27 amines used in constructing the model equation, which was 22.2% absolute average deviation (AAD). Validation of the degradation model was also carried out using MAE, AMP, and MEDA which were not used in generating the degradation model. The predicted degradation rates and experimental rates of the MAE, AMP, and MEDA were 2.13, 0.61, and 5.96 mM/h, respectively. The model could predict accuracies of 2.9, 10.3, and 7.8% AAD in comparison to the experimental rates of 2.07, 0.68, and 5.53 mM/h, respectively, for MAE, AMP, and MEDA, respectively. Although parameters such as amine bond length, spatial conformation, electronegativity inside and outside of the functional groups, and interactions between the functional groups could improve the linearity of the linear regression model, we opted to use a more advanced and comprehensive method based on machine learning to correlate the data without having to use the extra parameters since these parameters were not readily available in the industrial setting.

Table 2. Estimates of Degradation Model Parameters.

model parameter	coefficient	standard error (±)
LogK	2.908	0.474
a	0.049	0.012
b	0.132	0.010
c	0.176	0.017
f	–0.018	0.011
g	–0.078	0.015
h	0.087	0.017
i	0.056	0.014
l	0.021	0.010
n	2.533	0.364
o	–1.832	0.879

Figure 11.

Comparison of experimental and predicted degradation rates of 27 amines used in degradation rate model development using multiple linear regression and CatBoost machine learning.

Therefore, even though the multiple linear regression model could fairly predict the degradation rates of amines, the initially developed model predictability still needed further improvement because the assumption of linearity is not completely accurate. To achieve a better accuracy, a better statistical approach known as categorical boosting (CatBoost) was used to determine the relationship between the degradation rates and all 15 independent variables previously shown in Table 1. This refined approach integrated a machine learning technique to achieve a more comprehensive amine degradation predictive model. As in any regression analysis, all 27 amines were used to train the CatBoost model, and the resulting model produced 0.3% absolute averaged deviation (AAD), as seen in Figure 11. On the testing data set comprising MAE, AMP, and MEDA, the predicted degradation rates were 2.00, 0.71, and 5.60 mM/h, respectively, and the corresponding AAD were 2.9, 4.4, and 2.1%. Although the results from CatBoost are precise, there is no model to explain the statistical significance of the explanatory variable, as in the regression approach.

In order to confirm and establish that the degradation rates of the interpolated (synthetic) amines (SA-n) were authentic and reliable, their synthetic data points were respectively incorporated alongside the experimental amines with similarities in their chemical structures (i.e., chemical families) to assess degradation rate trends within these families. This is illustrated in Figure 12. Accordingly, SA-1 was categorized within the alkylenediamine family (Figure 12a), while SA-14 shared a classification similar to that of MEDA in the alkylenediamine family (Figure 12b). The alkanolamine family included SA-3, SA-4, and SA-6 (Figure 12c). SA-5 was placed in the sterically hindered alkanolamine family (Figure 12d), and SA-20 was associated with the alkyl cyclic-amine family (Figure 12e). Upon considering the interpolated degradation rates of the synthetic data, their values fell within the range and according to the trends established by the experimental ones. This comparative analysis confirmed and validated the data of the augmented data as being authentic and reliable.

Figure 12.

Examples of amine structures employed in the experimental amines alongside the augmented synthetic amines (the X-axis represents increasing C number).

Moreover, a comprehensive set of the 24 interpolated (synthetic) amine molecules, accompanied by their corresponding interpolated degradation rates and other relevant parameters, is presented in Table S3. A comparison of the results obtained from two mathematical models was applied to test amines (AMP, MAE, and MEDA), as depicted in Figure 13. The overall average absolute deviation (% AAD) for these three amines, as determined by the multiple linear regression model, was found to be 7.0%. However, the Catboost Model demonstrated a notable improvement, resulting in a reduced %AAD of 3.2%. These findings clearly indicate that the CatBoost model significantly enhances the accuracy of the degradation predictive model, as observed in both the training and test-amine sets. To validate the effectiveness and precision of the developed model, K-Fold Cross-Validation⁶⁰ was utilized to assess its robustness. The data set was divided into five nearly equal-sized subsets, and the model underwent training and evaluation five times. During each iteration, one subset was designated as the validation set, while the remaining four were employed for training. This process was repeated five times with a different subset designated as the validation set in each iteration. The mean MSE, calculated across these iterations, was 4.864125677955742 × 10^–7. The conspicuously low mean MSE value indicates the effectiveness and precision of our developed model in capturing the underlying patterns within the data.

Figure 13.

AAD comparison of three tested amines in the degradation predictive models developed by multiple linear regression and CatBoost machine learning methods.

4. Conclusions

Oxidative degradation in amines hinges on the stability of free radical formation, influenced by electronic and steric effects within the amine structure. Secondary amino groups are more prone to degradation due to their increased nucleophilicity, whereas cyclic amine structures resist degradation. Generally, longer alkyl chains between nucleophilic groups or at the nitrogen atom decrease degradation rates by reinforcing the N–C bond, thereby hindering free radical formation. This trend spans various amine categories. However, exceptions such as DAB and DETA with extended alkyl groups exhibit higher degradation rates due to the formation of stable cyclic structures that propagate degradation. The presence of hydroxyl groups accelerates amine degradation, particularly in secondary and tertiary alkanolamines. More hydroxyl groups can provide steric hindrance, reducing degradation rates in sterically hindered primary alkanolamines. A higher amino group abundance increases the oxidative degradation rates by offering more sites for free radical formation. Amines with branched alkyl groups between amino and hydroxyl groups or at the end of the alkyl group show increased degradation rates, especially when the branched alkyl group is between amino and hydroxyl groups.

The development of a semiempirical degradation rate model using multiple linear regression allowed for the prediction of amine degradation rates based solely on their structural groups and substituents, yielding an accuracy of 22.2% AAD. Through a validation test, the model demonstrated reasonable accuracy in predicting the degradation of unknown amines within the range of 7.0% AAD. The accuracy of prediction was drastically enhanced by employing the CatBoost machine learning approach. This approach yielded a degradation rate prediction accuracy of 0.3% AAD. Moreover, the AAD of the validation test was significantly reduced to 3.2%. Both models incorporated structural variables to account for the various chemical groups typically present in amines. Additionally, reaction and interaction variables were included in the model equation to consider the connectivity and reactivity effects of these structural groups, which can influence the amine degradation activity. The resulting model can be utilized by individuals to assess the degradation rate of a given amine before making a final decision on its suitability for use in a carbon capture plant.

Supporting Information Available

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsomega.3c07746.

• The Supporting Information is available for this article free of charge. Calculation examples for EWG, EDG, and S variables for model development; normalized degradation model variable values; comparison of the coefficient values obtained from original and normalized data sets; and comprehensive set of synthetic amine molecules, accompanied by the interpolated degradation rates and other relevant parameters (PDF)

The authors declare no competing financial interest.