A Comparative Study on the Carbonization of Chitin and Chitosan: Thermo-Kinetics, Thermodynamics and Artificial Neural Network Modeling

Abstract

The carbonization of chitin and chitosan presents a sustainable approach to producing nitrogen-doped carbon materials for various applications, making kinetic and thermodynamic analyses crucial for assessing their viability. Meanwhile, artificial neural network (ANN)-driven modeling not only enhances the precision of thermo-kinetic and thermodynamic estimations but also facilitates the optimization of carbonization conditions, thereby advancing the development of high-performance carbon materials. In this work, we aim to develop an ANN model to estimate weight loss as a function of temperature and heating rate during the carbonization of chitin and chitosan. The experimental average activation energies of chitosan and chitin, determined by various iso-conversional methods, were found to be 128.1–152.2 kJ/mol and 157.2–160.0 kJ/mol, respectively. The best-performing ANN architectures—NN4 for chitin (R² = 0.9995) and NN1 for chitosan (R² = 0.9997)—swiftly predicted activation energy values with commendable accuracy (R² > 0.92) without necessitating repetitive experiments. Furthermore, the estimation of thermodynamic parameters provided both a theoretical foundation and practical insights into the carbonization process of these biological macromolecules, while morpho-structural changes in the resulting chars were systematically examined across different carbonization temperatures. The results underscore the adaptability and effectiveness of ANN in analyzing the carbonization of biological macromolecules, establishing it as a reliable tool for thermochemical conversion studies.

1. Introduction

To reduce reliance on fossil fuels, there is a growing imperative to adopt a circular economy underpinned by bio-based materials and energy. This transition hinges on the sustainable sourcing of building blocks from non-fossil, abundant biogenic feedstocks—an approach that balances cost, environmental considerations, and resource availability. As a result, global interest in converting bio-based macromolecules into value-added chemicals, energy, and materials has intensified [1,2,3]. Among these candidates, chitin and chitosan stand out as promising natural polymers with distinct characteristics, poised to drive greener and more sustainable practices in various industrial and environmental applications.

Chitin, a poly-N-acetyl-D-glucosamine, is one of the most abundant, readily accessible, and renewable natural polymers—second only to cellulose. It is primarily derived from the exoskeletons of crustaceans such as crabs and shrimp. Depending on its source, chitin can take three polymorphic forms—α, β, and γ—each distinguished by variations in chain arrangement within their crystalline regions, resulting in distinct hydrogen bond networks. Of these, α-chitin is the most abundant and stable, featuring the most intricate hydrogen bond network [4,5]. On the other hand, chitosan, a deacetylated form of chitin, represents a polycationic polymer characterized by enhanced chemical versatility with a lower degree of acetylation [6]. It is defined as a linear copolymer featuring repeating units of d-glucosamine (deacetylated unit) and N-acetyl-d-glucosamine (acetylated unit), linked via β (1–4) glycosidic bonds [7]. Both chitin and chitosan are known to have favorable characteristics, including low cost, low toxicity, and widespread availability in addition to their excellent biocompatibility, biodegradability, and bioactivity. Particularly in the case of chitosan, its unique structure enables easier chemical modification [8]. However, from these compounds, only chitin is being isolated industrially. Chitin had a worldwide market valuation of approximately USD 1604.0 million in 2022 and is projected to achieve a value of USD 5026.4 million by the year 2032 [9]. Due to the inherent processability of chitin and chitosan, allowing their transformation into diverse formats such as sponges, gels, beads, scaffolds, micro and nanoparticles, these materials and their derivatives have been widely applied in various fields [10,11,12,13,14,15,16,17,18].

Carbonization (often referred to as pyrolysis) is an efficient and relatively straightforward strategy for transforming chitin and chitosan into value-added, nitrogen-doped carbon materials. Leveraging these biological macromolecules through well-designed carbonization processes is both desirable and feasible for producing functional carbon materials. A comprehensive understanding of carbonization and its associated decomposition mechanisms is pivotal for process design and optimization, as it offers in-depth insights into the breakdown pathways of these macromolecules [19]. This is because carbonization modeling employs mathematical frameworks to anticipate the yield and composition generated during thermal degradation. These models commonly integrate intricate chemical kinetics and thermodynamics to grasp the complex reaction pathways and rate dependencies [20].

In recent years, artificial neural network (ANN) approaches have contributed to modelling studies in various scientific and engineering domains, including pyrolysis and carbonization kinetics and thermodynamics [21,22]. Artificial neural networks (ANNs), inspired by the human brain, comprise layers of interconnected neurons that process and transform data [23,24]. Their ability to model complex data relationships makes them particularly useful in kinetic research [25]. In this study, ANN modeling was employed to validate and predict carbonization behavior thermodynamics of chitin and chitosan using thermogravimetric analysis (TGA). The primary goal of this study is to develop trained, ANN-based models that accurately predict kinetic parameters without multiple experimental trials. To achieve this, TGA data were collected during carbonization under various heating rates and a nitrogen atmosphere, and different functions were tested based on the ANN approach. Systematic parameter tuning then revealed the most effective ANN architectures for chitin and chitosan datasets. Next, kinetic and thermodynamic parameters were estimated using multiple models to assess the effectiveness of ANN in this context. The study’s novelty lies in applying ANN to both chitin and chitosan data and in successfully employing the optimized model to simulate kinetic and thermodynamic data. Further exploration was undertaken in order to examine morpho-structural changes during carbonization using various characterization tools, thereby offering deeper insights into the carbonization process and its underlying mechanisms.

2. Materials and Methods

2.1. Physicochemical Characterization and Thermal Analysis of Chitin and Chitosan

Commercial chitin and chitosan samples were obtained from Sigma-Aldrich (St. Louis, MO, USA) and used without any pretreatment. Prior to carbonization experiments, the samples were characterized using FT-IR spectroscopy (Perkin Elmer Spectrum 100, USA) and SEM-EDX (ZEISS Supra 40 VP, Germany), with the resulting spectra and micrographs presented in Figure 1 below.

The analyses revealed that both raw materials exhibit similar morphologies and contain carbon, oxygen, and nitrogen in their structures. Elemental composition analysis showed that chitosan contains 45.60 wt.% carbon, 10.61 wt.% hydrogen, 9.13 wt.% nitrogen, and 34.66 wt.% oxygen, whereas chitin has 49.88 wt.% carbon, 6.56 wt.% hydrogen, 8.69 wt.% nitrogen, and 34.87 wt.% oxygen. These results, further confirmed by an elemental analyzer (LECO CHN628), indicate that chitosan’s slightly higher nitrogen content arises from deacetylation during its derivation from chitin. In the FT-IR spectra of both chitin and chitosan, a broad band of around 3200–3400 cm⁻¹ corresponds to O–H stretching vibrations, and absorption bands at approximately 2850–2950 cm⁻¹ are attributed to C–H stretching in aliphatic structures. A C–O–C asymmetric stretching band near 1150 cm⁻¹ confirms the glycosidic linkages characteristic of polysaccharides. Key amide bands—indicative of the polymer’s backbone—include amide I (~1650 cm⁻¹, C=O stretching), amide II (~1550 cm⁻¹, N–H bending and C–N stretching), and amide III (~1300–1370 cm⁻¹, C–N stretching and N–H deformation). The relative intensity and position of these amide bands shed light on structural differences between chitin and chitosan, particularly regarding the degree of deacetylation.

Before performing TGA experiments on a Hitachi STA 7300 thermal analyzer, the crucibles and chamber were thoroughly cleaned with high-purity compressed air. Carbonization experiments were conducted under a nitrogen atmosphere at a constant carrier gas flow rate of 20 mL/min. To establish a baseline and reduce systematic errors, an empty alumina crucible was run under the same conditions before introducing the samples. A sample mass of 10 mg was maintained in each experiment to mitigate both mass and heat transfer limitations. Four different heating rates (5, 10, 20, and 40 °C/min) were used to investigate how heating rate influences thermochemical behavior in the temperature range of 25 °C to 1000 °C. All measurements were conducted in triplicate to ensure reproducibility, and consistent protocols were followed to enhance precision. Following TGA analysis, the samples were carbonized under the same conditions to obtain carbonaceous products at various temperatures. These chars were further characterized using SEM-EDX (ZEISS Supra 40 VP), Raman spectroscopy (Renishaw, Wotton-under-Edge, UK), and FT-IR spectroscopy (Perkin Elmer Spectrum 100) to evaluate morpho-structural changes during carbonization.

2.2. Kinetic Analysis

The fundamental rate equation for solid-state thermal decomposition processes typically assumes that the conversion rate is directly proportional to the reactant concentration and is temperature dependent. In non-isothermal experiments, the sample’s mass is tracked continuously as the reaction progresses under a constant heating rate, β (K/min). The kinetics of thermal degradation can be characterized by the rate of change of conversion, dα/dt. Under a linear heating rate, the kinetic expression is defined by two separate, independent functions—the temperature-dependent rate constant, k(T), and the fractional conversion function, f(α)—as shown in (1).

$${\displaystyle \frac{d\alpha }{dt}}=\beta {\displaystyle \frac{d\alpha }{dT}}=k(T)f(\alpha )$$

(1)

The dependency of the rate constant, k, on temperature is described by Arrhenius equation, where $E_a$ is the activation energy, $A$ is the pre-exponential factor and $R$ is the gas constant. as follows:

$$k(T)=\mathrm{Aexp}\left(-{\displaystyle \frac{E_a}{RT}}\right)$$

(2)

Integrating the temperature dependency of the reaction based on the Arrhenius law and subsequently adapting it for linear heating, (1) is transformed into (3), as follows:

$$\beta {\displaystyle \frac{d\alpha }{dT}}=A\exp \left(-{\displaystyle \frac{E_a}{RT}}\right)f(\alpha )$$

(3)

The conversion degree of the reaction is the extent of conversion and is represented by the following equation:

$$\alpha ={\displaystyle \frac{w_o-w_t}{w_o-w_f}}$$

(4)

where $w_o$ $w_t$ and $w_f$ represent the sample mass presented at initial, arbitrary, and final times respectively.

Equation (3) can also be integrated into

$${\displaystyle \underset{0}{\overset{\alpha }{\int }}\frac{d\alpha }{f(\alpha )}}=g(\alpha )={\displaystyle \frac{A}{\beta }}{\displaystyle \underset{To}{\overset{T}{\int }}\exp \left(-\frac{E_a}{RT}\right)}dT\equiv {\displaystyle \frac{A{E}_a}{\beta R}}p(u)$$

(5)

where g(α) and p(u) are known as the integrated form of fractional conversion function f(α) and the temperature integral, respectively. Then, various mathematical approximations can be employed to obtain the solution of p(u) through different approaches. In this study, four distinct models—Friedman, Flynn–Wall–Ozawa (FWO), Kissinger–Akahira–Sunose (KAS) and Starink —with iso-conversional approaches were utilized to compute the kinetic parameters.

Equations (6)–(9), below, depict the linearized representations of the kinetic models utilized in the iso-conversional kinetic analysis.

$$\ln \left(\beta {\displaystyle \frac{d\alpha }{dT}}\right)=\ln A+\ln f(\alpha )-{\displaystyle \frac{E_a}{RT}} \mathrm{(for\; Friedman)}$$

(6)

$$\ln \beta =\ln {\displaystyle \frac{A{E}_a}{Rg(\alpha )}}-5.331-1.052{\displaystyle \frac{E_a}{RT}}\hspace{1em} \mathrm{(for\; FWO)}$$

(7)

$$\ln \left({\displaystyle \frac{\beta }{T^2}}\right)=\ln \left({\displaystyle \frac{AR}{E_{a}g(\alpha )}}\right)-{\displaystyle \frac{E_a}{RT}}\hspace{1em}\hspace{1em} \mathrm{(for\; KAS)}$$

(8)

$$\ln \left({\displaystyle \frac{\beta }{T^{1.8}}}\right)=C_s-1.0037{\displaystyle \frac{E_a}{RT}}\hspace{1em}\hspace{1em}\hspace{1em}\mathrm{(for\; Starink)}$$

(9)

The requirement for achieving accurate and satisfactory values of activation energy is contingent upon the highest regression coefficient (R²) for the fitted regression line of the models.

2.3. Thermodynamic Analysis

Analogous to kinetic analysis, the pre-exponential factor, and variations in enthalpy (ΔH), Gibbs free energy (ΔG), and entropy (ΔS) can be computed utilizing thermogravimetric analysis, and these values are represented by (10)–(12).

$$A=\left[\beta E_a \exp \left({\displaystyle \frac{E_a}{R{T}_m}}\right)\right]/(R{T}_m^2)$$

(10)

$$\Delta H=E_a-RT$$

(11)

$$\Delta G=E_a+R{T}_m\ln \left({\displaystyle \frac{K_B{T}_m}{hA}}\right)$$

(12)

$$\Delta S={\displaystyle \frac{\Delta H-\Delta G}{T_m}}$$

(13)

where A is pre-exponential factor (s⁻¹), Tm is peak temperature (K), KB is the Boltzmann constant (1.381 × 10⁻²³ J/K) and h is the Plank constant (6.626 × 10⁻³⁴ Js).

2.4. Methodology of Artificial Neural Network Modeling

ANN can be described as a form of artificial intelligence within computing systems designed to process information in a manner analogous to the neural processing observed in the human brain [26]. The primary strength of ANNs lies in their ability to model complex, non-linear, and multidimensional functional relationships without requiring prior assumptions about these relationships. By leveraging self-organizing capabilities, they learn directly from empirical data, making them particularly well suited for capturing the behavior of nonlinear, multivariable systems—both static and dynamic [27].

The ANN approach, noted for its ability to learn from complex, nonlinear, and multidimensional data, is widely applied to prediction, classification, and optimization tasks. By generalizing from large training sets, ANNs can serve as “black-box” models for multivariable static and dynamic processes. Modeled after the human brain, a typical ANN comprises three layers—input, hidden, and output—each including a weight matrix, a bias vector, and an output vector that are able to capture intricate input–output relationships [28].

The transfer function of each layer in ANN, along with the interconnections between layers and neurons, plays a vital role in information processing. Commonly employed transfer functions in ANNs include linear (purelin), sigmoid or log-sigmoid, and hyperbolic tangent sigmoid functions. The dataset is typically partitioned into three subsets: training, validation, and testing. During training, the objective is to adjust the synaptic weights of the network to achieve a stable state. For model accuracy, the training data must be representative of the overall problem, as its characteristics critically influence the network’s learning performance [29,30].

In the proposed ANN model, the input parameters—heating rate (°C/min) and temperature (°C)—were employed to predict weight loss (%) behavior. This study utilized over 7000 lines of experimental TGA data. Following standard practice, 80% of the dataset was allocated for training, while the remaining 20% was reserved for validation. During training and validation, the dataset was split according to its features and via random selection, respectively. The ANN architecture, along with its corresponding input and output parameters, is illustrated in Figure 2. Further details regarding the ANN configuration used in this study are provided in Section 3.2.

In the estimation of thermodynamic parameters using ANN, activation energy values predicted through the ANN-assisted implementation of the Friedman method were employed as inputs for the corresponding thermodynamic equations. The experimentally determined thermodynamic functions were derived based on activation energy values obtained from the application of the Friedman method to non-isothermal TGA data. The Friedman method’s differential, model-free nature and high resolution make it well-suited for complex degradation systems, as its ability to yield accurate, conversion degree-specific activation energies enhance the reliability of thermodynamic parameter estimation and provides deeper insight into thermal degradation mechanisms.

3. Results and Discussion

3.1. Thermogravimetric Analysis

Figure 3, below, illustrates the TG thermograms for the carbonization of chitin and chitosan at heating rates ranging from 5 to 40 °C/min.

The key thermal parameters, including characteristic temperatures corresponding to the primary carbonization stages of each feedstock and differential TG (dTG), are also summarized in

Table 1. Characteristic temperatures for the carbonization of chitin and chitosan.

Sample	Heating Rate (°C/min)	Ti * (°C)	dTGmax **	Tf *** (°C)
Chitosan	5	218.7	287.8	366.2
	10	226.5	299.4	401.5
	20	238.1	310.5	504.3
	40	243.8	320.5	543.6
Chitin	5	233.7	392.4	434.8
	10	237.9	406.5	443.1
	20	241.2	423.9	529.1
	40	249.8	443.5	559.7

* Ti is the initial temperature of the main carbonization zone. ** dTGmax is the maximum temperature from the dTG curves. *** Tf is the final temperature of the main carbonization zone.

.

Chitosan generally exhibits a lower onset temperature (Ti) for the main carbonization stage compared with chitin at the same heating rates. This difference is primarily attributed to the deacetylation process in chitosan, which reduces the presence of acetyl (-COCH₃) groups and leads to a more amorphous structure with fewer intermolecular hydrogen bonds. Unlike chitin, which adopts a highly crystalline arrangement due to its acetyl content, chitosan’s amorphous nature enhances polymer chain mobility and increases accessibility to thermal energy, making it more susceptible to thermal degradation. The partial deacetylation of chitosan results in a higher concentration of amino groups (-NH₂) along the polymer chain, facilitating the disruption of both intermolecular and intramolecular hydrogen bonds. Compared with chitin, chitosan exhibits weaker intermolecular interactions, such as hydrogen bonding and van der Waals forces, due to its higher amino group content. These weaker interactions reduce the structural stability of the polymer, making the chains more prone to thermal decomposition at lower temperatures. Overall, the combination of partial deacetylation, an amorphous structure, and weaker intermolecular interactions in chitosan, relative to chitin, contributes to its lower onset temperature for the active degradation stage, as can be seen in the numerical results. Additionally, chitosan, which is partially deacetylated compared with chitin, generally exhibits a lower final temperature (Tf) for its main carbonization zone. During non-isothermal degradation under an inert atmosphere, the lower final temperature of active carbonization in chitosan is associated with deacetylation, increased amorphous structure, and a higher presence of reactive amino groups, all of which contribute to enhanced chain mobility.

As the heating rate increased from 5 to 40 °C/min, distinct shifts of the thermograms of both chitin and chitosan toward higher temperature regions were observed, demonstrating a well-defined and predictable thermal hysteresis. This phenomenon is primarily attributed to the relatively low thermal conductivities of materials degraded in thermogravimetry. To put it in other words, the temperature gradient between the surface and the interior of the feedstock particles was reduced, enabling more uniform heating and facilitating decomposition due to an extended residence time at lower heating rate values. Furthermore, the peak temperatures obtained from dTG profiles exhibited an increase in peak intensities together with increasing peak temperatures particularly in the maximum weight loss rates at each stage, where the variations were most pronounced. The peak temperatures (dTGmax) represent the temperature at which the maximum rate of mass loss occurs during thermal degradation. The dTG peak temperature gradually increased from 287.8 to 320.5 °C for chitosan and from 392.4 to 443.5 °C for chitin, indicating thermal lag during carbonization.

The residual weight left after the degradation at 1000 °C for chitosan and chitin were found to be in the range of 32.6–34.1 wt.% and 10.5–10.9 wt. %, respectively, at various heating rates. Clearly, chitosan yielded a higher char fraction than chitin when thermally degraded in a nitrogen atmosphere, primarily due to differences in chemical composition, molecular structure, and thermal degradation mechanisms. Chitosan, possessing a higher nitrogen content due to its increased number of amino (-NH₂) groups compared with chitin, which contains more N-acetyl (-NHCOCH₃) groups, facilitates the formation of cross-linked nitrogen-containing structures that enhance the stability of the carbonaceous char. The presence of nitrogen in the polymer backbone also promotes cyclization reactions and the development of thermally stable aromatic structures, thereby reducing the extent of volatile release while increasing char yield during carbonization. This can be explained by the presence of nitrogen in the polymer backbone, which favors cyclization reactions and the formation of thermally stable aromatic structures, reducing the extent of volatile release. Due to its higher degree of acetylation, chitin contains a greater proportion of N-acetylglucosamine units, which primarily decompose through extensive deacetylation and dehydration, leading to increased volatile release compared with chitosan. In contrast, chitosan, being partially deacetylated, undergoes carbonization with reduced deacetylation-driven mass loss, resulting in a higher retention of solid residue. Ultimately, chitosan’s higher nitrogen content, lower degree of acetylation, enhanced cross-linking, and reduced volatile release collectively contribute to its higher char yield compared with chitin under identical carbonization conditions. Additionally, the formation of nitrogen-enriched aromatic structures reinforces the stabilization of the carbonaceous matrix, further promoting char retention in chitosan.

3.2. Computational Results of ANN

In research on the thermal degradation process using the ANN approach, the data are typically partitioned randomly into three subsets: training, validation, and testing. However, the primary goal of this study was to reduce the number of experimental trials by employing the ANN model. To achieve this, various combinations of heating rates—selecting three rates from four distinct values—were utilized during training to predict the remaining one. To demonstrate the applicability of ANN models to real-life scenarios, the dataset was divided into two parts for the training and testing phases. These two parts consist of data with different heating rates so that predicting a heating rate not used during the training phase would highlight the realistic capabilities of the ANN models. This procedure helps to minimize the risk of overfitting in the studied data. For example, if the 5, 10, and 20 °C/min heating rates were selected for training, the model would be tested to predict the remaining heating rate, which is 40 °C/min. In this study, 20% of the datasets were split away indiscriminately for the validation process. The ANN model was implemented in Python using the Keras library with TensorFlow as the backend. The Adam optimizer was employed for backpropagation. Two different transfer functions, Tansig and Sigmoid, were tested. The network models were trained for 1000 iterations with a target error of 10⁻⁶. Multiple runs were performed with different parameters to compare the neural models. Various architectures with different numbers of layers and neurons were tested across these runs. The parameters used and the corresponding results for two distinct datasets—chitin and chitosan experiment sets—are presented in

Table 2. ANN models tested for the chitin and chitosan data and the regression performance results.

					Chitin Data		Chitosan Data
Models	Network Structure	Heating Rate for Training	Heating Rate for Test	Transfer Function	Test R2 Score	Validation R2 Score	Test R2 Score	Validation R2 Score
NN1	5-10-5	5 °C/min 10 °C/min 20 °C/min	40 °C/min	Tansig	0.9789	0.9899	0.9997	0.9999
NN2	5-10-10	5 °C/min 10 °C/min 40 °C/min	20 °C/min	Sigmoid	0.97569	0.9995	0.97569	0.9995
NN3	5-20-5	5 °C/min 20 °C/min 40 °C/min	10 °C/min	Tansig	0.9648	0.9799	0.9854	0.9899
NN4	10-20-10	10 °C/min 20 °C/min 40 °C/min	5 °C/min	Sigmoid	0.9995	0.9999	0.9648	0.9799
NN5	5-10-20-10	5 °C/min 20 °C/min 40 °C/min	10 °C/min	Tansig	0.756	0.7579	0.9125	0.9156
NN6	5-15-30-15	10 °C/min 20 °C/min 40 °C/min	5 °C/min	Tansig	0.8328	0.855	0.9365	0.9499
NN7	5-20-5-10	5 °C/min 10 °C/min 40 °C/min	20 °C/min	Sigmoid	0.8966	0.9122	0.9256	0.9269
NN8	1-5-1	5 °C/min 20 °C/min 40 °C/min	10 °C/min	Tansig	0.6242	0.6658	0.7458	0.7556
NN9	2-3-2	5 °C/min 10 °C/min 20 °C/min	40 °C/min	Sigmoid	0.6899	0.7156	0.7025	0.6987
NN10	5–20-10-20-5	5 °C/min 20 °C/min 40 °C/min	10 °C/min	Tansig	0.9723	0.9756	0.9568	0.96667
NN11	1-5-1	5 °C/min 10 °C/min 20 °C/min	40 °C/min	Sigmoid	0.8896	0.8936	0.7895	0.8026
NN12	5–20-30-20-5	5 °C/min 10 °C/min 40 °C/min	20 °C/min	Sigmoid	0.9883	0.9901	0.9366	0.93999
NN13	3-10-2	10 °C/min 20 °C/min 40 °C/min	5 °C/min	Sigmoid	0.899	0.9199	0.9124	0.9205
NN14	1-3-1	10 °C/min 20 °C/min 40 °C/min	5 °C/min	Tansig	0.7455	0.7041	0.7887	0.7999
NN15	1-3-1	5 °C/min 10 °C/min 20 °C/min	40 °C/min	Tansig	0.7856	0.7745	0.6698	0.6784

, which summarizes the parameters used and the results obtained for both chitin and chitosan data.

Table 2 also provides a detailed summary of the unique neural network structures (e.g., 5-10-5 indicates three hidden layers with 5, 10, and 5 neurons, respectively), along with the details of the corresponding neural architecture and the combination of heating rate data used during the training, validation, and testing processes. Additionally, the transfer functions utilized in the model, as well as the test and validation regression coefficient (R²) scores for chitin and chitosan, are presented in the above table. To evaluate performance, the R² score was calculated. Generally, R² values are assessed in three ranges: R² > 0.9 shows desirable performance and R² values between 0.8 and 0.9 indicates fair performance, while R² < 0.8 implies unsatisfying performance of the model. According to the results, the best-performing neural network model structure for the chitin data is NN 4, which consists of three layers (10-20-10) with a Sigmoid transfer function, achieving an R² value of 0.9995. For the chitosan data, the best-performing model is NN 1, which also has three layers (5-10-5) but uses a Tansig transfer function, achieving an R² value of 0.9997. As the underlying mechanisms of the thermal degradation of chitin and chitosan differ, this results in certain mathematical complexities during model development. Consequently, the same neural structure may yield different performance results across different datasets.

Given that the regression model of the TGA is a complex polynomial function with a relatively small number of parameters, the ANN model requires increased complexity to achieve accurate predictions. To evaluate model performance, mean squared error (MSE), root mean square error (RMSE), and mean absolute error (MAE) were used as key indicators. The TG data were normalized to a 0–1 scale to clearly emphasize the performance outcomes. Among the tested models, only those demonstrating stability and high performance were selected for further analysis. The detailed statistical results are presented in

Table 3. Comparison of the statistical parameters of different ANN models.

Model Structure	Number of Layers	Total Number of Neurons	Complexity Level	Chitosan			Chitin
				MSE	RMSE	MAE	RMSE	MAE	MSE
1-3-1	3	5	1	0.0012	0.0326	0.0311	0.0233	0.0211	0.0018
2-4-2	3	8	2	0.009	0.0254	0.0255	0.0189	0.0375	0.007
5-5-5	3	15	3	0.0010	0.0266	0.0199	0.0298	0.0149	0.008
5-10-5	3	20	4	0.0016	0.0189	0.0143	0.0156	0.0259	0.0010
5-20-5	3	30	5	0.0015	0.0195	0.0281	0.0181	0.0247	0.0012
10-20-10	3	40	6	0.009	0.0233	0.0131	0.0259	0.0254	0.0015
5-10-20-5	4	40	7	0.0012	0.0278	0.0125	0.0222	0.0355	0.0010
5-20-20-5	4	50	8	0.0017	0.0398	0.0081	0.0269	0.0092	0.0014
10-20-20-10	4	60	9	0.0012	0.0415	0.0237	0.0369	0.0287	0.0011
15-20-20-15	4	70	10	0.009	0.0429	0.0393	0.0355	0.0375	0.007
10-20-20-10-10	5	70	11	0.0018	0.0548	0.0449	0.0487	0.0426	0.0015
15-20-20-20-15	5	90	12	0.0020	0.0619	0.0505	0.0512	0.0445	0.0019

.

The table above presents the performance metrics of ANN models with varying architectures for the chitosan and chitin datasets. It lists 12 different ANN architectures with increasing complexity, ranging from level 1 to level 12, representing growth in the number of layers and total neurons. The simplest structure is 1-3-1, comprising just 5 neurons across 3 layers, while the most complex is 15-20-20-20-15, with 90 neurons distributed across 5 layers. The minimum MSE value for chitosan was calculated for the model structure 5-5-5 at complexity level 3 with 15 neurons. For chitin, the minimum MSE value was found in the model structure 5-10-5 at complexity level 3 with 20 neurons.

Figure 4a,b, below, compare the experimental data and model predictions for various heating rates in the chitosan and chitin datasets, respectively. In these figures, the x-axis represents the actual heating rate values, while the y-axis represents the predicted heating rate values. The results reveal a nearly perfect linear correlation between the actual values (blue points) and the predicted values (red points), indicating minimal prediction error [31]. The predicted values almost perfectly overlap with the actual data, demonstrating a well-trained model with strong generalization capabilities—especially if these results are obtained from a test set. The analysis results underscore the power of ANNs in predicting outcomes in thermal degradation experiments. It is evident that further weight loss data with different parameters can be realistically predicted using the experimental results obtained through ANN modeling.

3.3. Kinetics

Figure 5 presents a comparative analysis of the variation in activation energy as a function of the degree of conversion for both chitin and chitosan. Additionally, different kinetic models have been evaluated and compared. For this purpose, the Friedman, Flynn–Wall–Ozawa (FWO), Kissinger–Akahira–Sunose (KAS), and Starink iso-conversional methods were analyzed. The kinetic parameter calculation for the carbonization of chitin and chitosan was calculated assuming a reaction order of unity (using the fractional conversion function f(α) = (1 − α)ⁿ, as suggested by White et al. [32]. This approach is supported by findings that the thermal degradation of these macromolecules closely follows an apparent first-order reaction model, which effectively approximates their decomposition kinetics [33]. Specifically, the reaction order being close to unity is consistent with previous studies, which have shown that a first-order kinetic model sufficiently describes the thermal degradation behavior of both chitin and chitosan. To calculate the pre-exponential factor (A), the Friedman method was employed due to its ability to provide accurate estimates through a direct differential approach, correlating the reaction rate with temperature at constant conversion levels. This method complements the activation energy determination from iso-conversional analysis by avoiding integral approximations, thereby yielding reliable kinetic parameters even for complex degradation processes. Figure 5 demonstrates that the Friedman, FWO, KAS, and Starink models exhibit a similar trend in activation energy values across conversion degrees ranging from 0.1 to 0.9. This means that the thermal degradation process follows a similar pathway, regardless of the specific kinetic approach used. To put it in other words, the similar trend in activation energy estimation across different methodologies suggests that the material exhibits intrinsic kinetic behavior that is accurately captured by all four models. However, variations in numerical values were observed across the utilized models, attributable to differences in their underlying assumptions and approximations.

The average activation energy of the carbonization process provides insight into the thermal degradation behavior, where a lower activation energy indicates that degradation occurs more readily and at lower temperatures in chemical kinetics. The average experimental activation energies for the carbonization of chitosan, as determined using the Friedman, FWO, KAS, and Starink methods, were 152.2 kJ/mol, 131.1 kJ/mol, 128.1 kJ/mol, and 128.6 kJ/mol, respectively. While the average experimental activation energies in chitin carbonization for the Friedman, FWO, KAS, and Starink methods were 160.0, 159.9, 157.3, and 157.8. The higher degree of acetylation in chitin, characterized by a greater proportion of N-acetylglucosamine units, enhances intermolecular interactions, leading to increased structural stability and a higher activation energy requirement for carbonization. In contrast, chitosan undergoes partial deacetylation, resulting in a reduced number of acetyl groups. This reduction weakens hydrogen bonding and van der Waals interactions compared with chitin, thereby lowering the activation energy needed for carbonization. Consequently, chitosan exhibits a lower average activation energy for carbonization than chitin, indicating that it undergoes thermal degradation more readily at lower temperatures. In addition, fluctuations in activation with respect to the conversion degree for both chitosan and chitin were observed. This suggests that the carbonization of chitosan and chitin is not governed by a single uniform reaction but involves a sequence of complex, temperature-dependent transformations. Different functional groups in the structure may degrade at different temperatures, leading to shifts in activation energy, while processes such as aromatization, dehydrogenation, or intermediate char formation can further introduce fluctuations in the apparent activation energy for the carbonization of chitosan and chitin.

In the final stage of the primary thermal decomposition zone, the activation energy of chitosan exhibited a decreasing trend, whereas for chitin, it showed an increasing tendency. The divergence in activation energy trends during the final stage of the primary thermal decomposition zone can be attributed to differences in their chemical structures and thermal degradation pathways. In the final stage, chitosan undergoes extensive char formation and depolymerization, with residual polymer breakdown occurring at lower energy barriers. The presence of amino groups facilitates fragmentation, reducing activation energy, while secondary reactions, such as volatile release and char rearrangement, further contribute to this decline. On the other hand, the increasing activation energy in the final stage indicates dominant cross-linking and intermolecular interactions for chitin. Residual acetyl groups in chitin enhance char stabilization and aromatization, strengthening bonds and requiring higher energy for degradation. Additionally, condensed carbonaceous structures may introduce diffusion limitations, further raising activation energy for carbonization of chitin at the last stages of thermal degradation.

The activation energy results also indicate that the activation energy trends closely align with predictions obtained using artificial neural networks. Furthermore, the findings demonstrate that artificial neural networks can accurately predict the average activation energy value. The percentage error between the experimental and predicted average activation energy values ranged from 0.11% to 2.88%, confirming the reliability of the predictive model. The lowest error is observed for the prediction of the chitosan carbonization using the KAS (0.20%) and FWO (0.11%) models, suggesting that these models provide the most accurate predictions for chitosan. Similarly, for chitin, the KAS model (0.12%) demonstrates the smallest deviation, indicating high reliability in prediction. The average predicted activation energies for the carbonization of chitosan, as determined using the Friedman, FWO, KAS, and Starink methods, were 156.5 kJ/mol, 131.3 kJ/mol, 128.4 kJ/mol, and 128.8 kJ/mol, respectively. On the other hand, the average experimental activation energies for chitin carbonization, as determined using the Friedman, FWO, KAS, and Starink methods, were 158.4, 162.4, 157.1, and 160.3 kJ/mol, respectively. Overall, FWO and KAS models appear to provide the most consistent and accurate predictions for both chitosan and chitin. While the Friedman and Starink models also perform well, they show slightly higher deviations in some cases, indicating potential limitations depending on the material. The small percentage errors suggest that all models can be considered reliable for predicting the behavior of chitosan and chitin, but that KAS and FWO are preferable choices for minimal deviation.

Table 4. Changes in the pre-exponential factor as a function of conversion rate and compensation effect.

		Experimental		Prediction
	α	A (s−1)	Compensation Plot Equation	A (s−1)	Compensation Plot Equation
Chitosan	0.1	1.53 × 1012	lnA = 0.2173Ea − 8.6225	6.81 × 1012	lnA = 2172Ea − 8.6148
	0.2	4.55 × 1013		2.90 × 1015
	0.3	8.66 × 1012		3.54 × 1012
	0.4	5.79 × 1013		4.87 × 1013
	0.5	1.37 × 1014		3.67 × 1012
	0.6	1.79 × 109		3.32 × 1012
	0.7	7.47 × 108		5.17 × 108
	0.8	2.96 × 105		3.30 × 105
	0.9	1.83 × 105		2.50 × 105
	$\overline{\mathrm{x}}$	2.78 × 1013		3.29 × 1014
Chitin	0.1	5.41 × 109	lnA = 0.1836Ea − 8.7791	4.82 × 109	lnA = 0.1836Ea − 8.7711
	0.2	3.39 × 108		2.56 × 108
	0.3	8.81 × 1010		1.12 × 1010
	0.4	1.39 × 107		9.81 × 106
	0.5	1.98 × 107		1.17 × 107
	0.6	3.56 × 107		2.31 × 107
	0.7	7.83 × 107		1.13 × 108
	0.8	2.48 × 1010		7.19 × 109
	0.9	1.09 × 1011		9.89 × 1011
	$\overline{\mathrm{x}}$	2.54 × 1010		1.13 × 1011

shows the variation of the pre-exponential factor during the carbonization processes of chitin and chitosan.

The significant variation in the pre-exponential factor as a function of the conversion degree reflected intricate reactions occurring during thermal degradation. As the pre-exponential factor represents the frequency of collisions between reactant molecules, variations in its value with changing conversion degrees can be attributed to differences in reaction chemistry and complex formation. A low pre-exponential factor (<10⁹ s⁻¹) is associated with surface reactions, whereas in reactions independent of surface area, a low value suggests the presence of a tightly bound, closed complex. Conversely, a high pre-exponential factor (≥10⁹ s⁻¹) indicates a loosely bound, junctional complex. Additionally, higher pre-exponential factors reflect greater sensitivity to temperature variations within the reaction range [34]. For chitosan, pre-exponential factor values were found to be higher than 10⁹ s⁻¹, up to a conversion degree of 0.7. This signifies a transition to a loosely bound, high-mobility reaction state and the increased influence of temperature on reaction kinetics. As carbonization progressed, the structure of chitosan has been broken down into smaller, less reactive fragments, leading to a reduction in the frequency of effective molecular collisions, which lowers the pre-exponential factor. The progressive formation of a thermally resistant char matrix may hinder further decomposition, reducing the number of available reaction sites and consequently lowering the pre-exponential factor to approximately 10⁵ s⁻¹ in the final stage of chitosan carbonization. On the other hand, chitin possessed a highly ordered crystalline structure due to extensive hydrogen bonding, which requires a higher frequency of molecular collisions to sustain decomposition reactions, leading to larger pre-exponential values (≥10¹⁰ s⁻¹) compared with chitosan at the last stage of thermal decomposition. The linear correlation between lnA and Ea, which facilitates the carbonization reactions in terms of compensation effect, is also depicted in Table 4. The regression analysis for both chitin and chitosan carbonization produced high regression coefficients, confirming a strong linear relationship among each other. In other words, the activation energy (Ea) and pre-exponential factor (A) exhibit a near-perfect linear dependence, which can be described using an lnA = aEa + b mathematical model, supporting the concept of energy compensation. The estimated equations with ANN were also closely matched the experimentally obtained expressions with high regression coefficients. This consistency suggests that the ANN models employed in this study demonstrate strong kinetic prediction capability and generalization performance in modeling the relationship between activation energy and the pre-exponential factor for both chitosan and chitin carbonization.

3.4. Thermodynamics

The enthalpy change, entropy change, and Gibbs free energy change values predicted using artificial neural networks (ANNs) were calculated using the experimentally obtained data and the activation energy derived from the Friedman method. The results are presented in Figure 6 and

Table 5. Average values of the thermodynamic parameters for carbonization.

		∆H (kJ/mol)	∆G (kJ/mol)	∆S (J/mol)
Chitosan	Experimental	155.2	168.2	−22.7
	Prediction	151.7	166.3	−25.5
	% Deviation	2.2	1.1	12.3
Chitin	Experimental	154.6	183.7	−42.9
	Prediction	153.1	182.9	−43.9
	% Deviation	0.9	0.4	2.3

in detail.

The experimental average enthalpy change (ΔH) values for chitosan and chitin were 155.2 kJ/mol and 154.6 kJ/mol, respectively. The predicted average ΔH values were slightly lower than the experimental values (151.7 kJ/mol for chitosan and 153.1 kJ/mol for chitin), but the percentage deviation remained low, which indicates a good agreement between model predictions and actual values. These values confirm that the carbonization of chitosan and chitin is an endothermic process, requiring significant thermal energy for thermal decomposition. The average experimental Gibbs free energy change (ΔG) values were found to be significantly higher for chitin (183.7 kJ/mol) compared with chitosan (168.2 kJ/mol), indicating that chitin carbonization is thermodynamically less favorable during the main thermal decomposition stage. The higher average ΔG for chitin suggests a more stable structure, requiring greater energy to reach the transition state, possibly due to its higher crystallinity and stronger interchain hydrogen bonding. The average entropy change (ΔS) values were negative for both materials, indicating a decrease in system disorder during carbonization, likely due to the formation of more ordered char structures. Chitin showed a much more negative entropy change (−42.9 J/mol·K vs. −22.7 J/mol·K for chitosan), suggesting a greater degree of structural reorganization and condensation during thermal degradation. Overall, chitin exhibited a higher average ΔG and more negative average ΔS, indicating greater thermal stability and a more structured char formation process compared with chitosan. On the other hand, chitosan’s lower average ΔG and less negative average ΔS suggests a more favourable and less structured decomposition pathway, which may be due to its higher degree of deacetylation and lower crystallinity.

When Figure 6 was carefully investigated to determine instantaneous changes in the thermodynamic parameters, ANN predictions for chitin demonstrated better accuracy in instantaneous estimations when compared with those for chitosan. These findings indicate that the average thermodynamic parameters can be accurately predicted; however, deviations in specific thermodynamic parameters may occur using ANN. The ANN learned trends for average values of the changes in thermodynamic functions effectively but struggled with extreme values or cases. Computational approximations in ANN training, such as overfitting, data bias, or noise in the input variables, may contribute to minor deviations in specific thermodynamic values at the specific conversion degree point.

3.5. Morpho-Structural Changes During Carbonization

To gain a more comprehensive understanding of the structural changes occurring during the carbonization of chitin and chitosan, carbonized chars at different temperature levels were obtained at the same conditions and the characteristics of the chars were investigated using FT-IR and Raman spectroscopic techniques. Moreover, SEM-EDX analyses were performed to identify the structural and chemical changes occurring during carbonization, with the solid products obtained at different temperatures.

Figure 7 shows Raman and FT-IR spectra of the carbonized products of chitin and chitosan at various temperatures. In the Raman spectrum, one can see the G-band (~1580 cm⁻¹) and D-band (~1350 cm⁻¹), and their intensity ratios (ID/IG) can be used to assess the structural properties and disorder levels in carbonaceous materials. For both chitosan and chitin, the ID/IG ratio was increased with increasing carbonization temperatures. The ratios for chitosan were found to be 0.863, 0.938, and 0.977 at carbonization temperatures of 500 °C, 700 °C, and 900 °C, respectively. On the other hand, for chitin, the ratios were 0.921, 0.923, and 0.948 at the same respective temperatures. At higher carbonization temperatures, chitosan-derived chars exhibit increasing structural disorder, as evidenced by the rising ID/IG ratio. This suggests that thermal treatment induces bond rearrangements, leading to the fragmentation of smaller graphitic domains and an increase in defect density. Consequently, chitosan chars are more prone to forming amorphous or turbostratic carbon structures rather than well-ordered graphitic domains. In contrast, chitin-derived chars demonstrate greater structural stability across different carbonization temperatures, with a relatively stable ID/IG ratio. This indicates a lower degree of structural transformation compared with chitosan. The increased disorder in chitosan chars results from differences in molecular structure and thermal decomposition mechanisms compared with chitin, affecting defect generation during carbonization. This can be attributed to chitin’s inherently more rigid and highly ordered structure, reinforced by strong intra- and intermolecular hydrogen bonding, whereas chitosan, with fewer acetyl groups, follows a distinct thermal decomposition pathway that influences its structural evolution during carbonization. Raman spectroscopy further confirms that neither chitosan nor chitin undergoes full graphitization, as indicated by the relatively high intensity ratio values, which suggest the persistence of a predominantly amorphous or turbostratic carbon structure.

From the FT-IR spectra, O-H (hydroxyl) and N-H (amine) stretching is observed at 3200–3500 cm⁻¹, as well as C=O (carbonyl, amide I) stretching. This band decreased with increasing temperature due to dehydration and deamination. Additionally, C=O (carbonyl, amide I) stretching bands around 1650–1700 cm⁻¹ were found to be strong at 500 °C, then weakened due to amide degradation. N-H bending (amide II) and C=C stretching in aromatic structures, at around 1550–1600 cm⁻¹, were shifted to higher temperatures due to aromatization. Meanwhile, the C-H bending stretching around 1400–1450 cm⁻¹ decreased as aliphatic structures degraded during carbonization, while aromatic C-H bending was observed at around 800–900 cm⁻¹, indicating the formation of polyaromatic structures.

SEM-EDX analyses are also shown in Figure S1 to investigate the thermal degradation effects on the chars. Morphologically, and with increasing temperature, localized pulverization, increased surface roughness, and fractures were observed in the structure, while the elemental carbon content rose to 92.2% for chitosan and 89.3% for chitin. As the polysaccharide chains fragmented due to carbonization, the original structure of chitin and chitosan collapsed, leading to the loss of molecular order. With rising temperature, thermal degradation further broke down the polysaccharide chains into smaller molecules, facilitating the release of volatile components such as hydrogen and oxygen. Consequently, the remaining solid product became enriched in carbon. Ultimately, the carbonization process transformed chitin and chitosan into carbon-rich solid products with a significantly higher elemental carbon content compared with the original materials.

4. Conclusions

ANN models, which offer valuable insights into the kinetics of complex, nonlinear systems, were found to yield predicted kinetic parameters that closely align with experimentally determined values, indicating strong predictive performance when properly optimized and well trained. Notably, the models accurately captured trends in activation energy, pre-exponential factors, the kinetic compensation effect, and thermodynamic parameters—including enthalpy, entropy, and Gibbs free energy. Furthermore, the Friedman, FWO, KAS, and Starink models displayed similar activation energy trends across conversion degrees, reinforcing the notion that the carbonization of chitosan and chitin follows a consistent kinetic pathway that is effectively mirrored by ANN predictions. Chitosan’s average activation energy ranged from 128.1 to 152.2 kJ/mol, while chitin’s was higher (157.3 to 160.0 kJ/mol) due to stronger intermolecular interactions from higher acetylation. In the final stage, chitosan’s activation energy decreased due to char formation and polymer breakdown, whereas chitin’s increased due to cross-linking and char stabilization. Fluctuations in activation energy suggest a complex, temperature-dependent degradation process for both materials. Experimental ΔH values for chitosan (155.2 kJ/mol) and chitin (154.6 kJ/mol) confirmed the endothermic nature of carbonization, with ANN predictions closely aligned but slightly lower. Chitin exhibited higher ΔG (183.7 kJ/mol) than chitosan (168.2 kJ/mol), indicating greater thermal stability due to its crystallinity and strong hydrogen bonding. Both materials showed negative ΔS, with chitin’s more negative value (−42.9 J/mol·K vs. −22.7 J/mol·K) suggesting greater structural reorganization. ANN predictions accurately captured average thermodynamic trends but showed minor deviations in instantaneous values. Morpho-structural analysis of carbonaceous products at different carbonization temperatures revealed that surface morphology, degree of disorder, and surface functionality of the resulting chars were significantly influenced by both the precursor material and the carbonization temperature. The results of this study confirm that the integrated experimental methodology—combining thermo-kinetic analysis, thermodynamics and ANN modeling—is effective in providing valuable insights into the kinetic modeling of complex nonlinear systems. Consequently, the predictive capabilities of the ANN approach in both kinetics and thermodynamics contribute to a deeper understanding of the carbonization processes of chitin and chitosan. This improved understanding facilitates precise control over product characteristics, supports efficient scale-up and process optimization, and enhances resource utilization, ultimately enabling more cost-effective processing.