Thermal degradation model of used surgical masks based on machine learning methodology

Abstract

Background

The COVID-19 pandemic has leveraged facial masks to be one of the most effective measures to prevent the spread of the virus, which thereby has exponentially increased the usage of facial masks that lead to medical waste mismanagements which pose a serious threat to life. Thermal degradation or pyrolysis is an effective treatment method for the used facial mask wastes and this study aims to investigate the thermal degradation of the same.

Methods

Predicted the TGA experimental curves of the mask components using a Machine Learning model known as Artificial Neural Network (ANN).

Significant findings

Three different parts of the mask namely- ribbon, body, and corner were separated and used for the analysis. The thermal degradation behavior is studied using Thermogravimetric Analysis (TGA) and this is crucial for determining the reactivity of the individual mask components as they are subjected to a range of temperatures. Using the curves obtained from TGA, kinetic parameters such as Activation energy (E) and Pre-exponential factor (A) were estimated using the Coats-Redfern model-fitting method. Using the determined kinetic parameters, thermodynamic quantities such as a change in Enthalpy (ΔH), Entropy (ΔS), and Gibbs-Free energy (ΔG) were also calculated. Since TGA is a costly and time-consuming process, this study attempted to predict the TGA experimental curves of the mask components using a Machine Learning model known as Artificial Neural Network (ANN). The dataset obtained at a heating rate of 10°C/min was used to train the 3 different neural networks corresponding to the mask components and it showed an excellent agreement with experimental data (R² > 0.99). Through this study, a complex chemical process such as thermal degradation was modelled using Machine Learning based on available experimental parameters without delving into the intricacies and complexities of the process.

Graphical abstract

1. Introduction

The COVID-19 pandemic, ever since its declaration by the World Health Organization (WHO), has completely altered our daily lives, work styles, and activities. With the resurgence and downfall of infections, there is not a precise answer to when this global pandemic shall end, as a result of which people and communities are well aware of the various ways of protecting themselves from getting infected. Out of all the ways suggested by experts, the usage of masks has proven to be very effective to avoid getting infected by the virus. It has been estimated from lab experiments that, using facial masks can block nearly 50–70% of exhaled small droplets and particles which are the prime carriers of the virus [1]. The pandemic has led to increased awareness about the usage of face masks as protective gear and has led to an increase sales which are well supported by the fact that in 2019 the global facial mask market was valued at 6792 million USD and is estimated to reach 9052 million USD by 2027. The local governments have been raising awareness about using facial masks when going out in public places, this has caused an exponential rise in the usage of masks ever since. Attributed to the non-reusability of masks, the waste generation and disposal methods have raised concerns for the local municipality. In 2019, nearly 400 MT of plastic waste is produced globally but within the first 7 months since the outbreak of the virus, the volume of waste reached over 500 MT [2]. Additionally, in a press conference study of the Joint Prevention and Control Mechanism of the State Council of China, it was found that nearly 468.9 tons of medical wastes associated with COVID-19 are generated per day [3].

The COVID-19 pandemic scenario has raised concerns about the adversities of improper handling of medical wastes which if not done properly can lead to the spread of diseases. The mismanagement of millions of tons of waste comprising used PPE equipment such as masks and gloves has not only posed serious health and environmental threat but also has levitated the spread of the Coronavirus more quickly [4]. Because these masks are made of plastics such as polystyrene, polycarbonate, polypropylene, polyurethane, polyacrylonitrile, and polyethylene, they contribute to micro plastic pollution and thus require sustainable disposal methods. Fig. 1 depicts the various waste disposal techniques, the selection of the appropriate waste disposal strategy might vary from region to region in compliance with environmental and local authority regulations.

Fig. 1.

Waste disposal technologies for biomedical wastes.

Among all the disposal technologies, this study primarily focuses on the thermal treatment such as pyrolysis of used surgical mask wastes. This is an energy-saving method that allows a complete decomposition of biomedical waste and depending on temperature regulations it offers destruction of toxins such as dioxins and furans. In pyrolysis, various plastic is mixed and converted to liquid fuels for energy generation to power various industrial applications [5], [6], [7]. Another additional benefit pertaining to the pyrolysis method is that it allows for the reduction of CO₂ and other toxic gasses alongside which it contributes to the production of valuable fuels such as syngas and liquid fuels which can be used for domestic and industrial applications [8]. Fig. 2 shows the different parts of a surgical mask, every part is made up of different polymer material that reacts with heat differently across various temperature regimes. In the literature, various studies have been conducted to determine the chemical composition of the mask and different methods of characterization have been adopted, such as Fourier Transform Infrared Spectra (FTIR) [[9], [10], [11]], Pyrolysis Combustion Flow Calorimetry (PCFC), and Mass Spectroscopy (MS) [[12], [13], [14]]. Based upon inference from the results, a commonly used surgical mask is made up of polymeric materials such as polypropylene, whilst they are also made up of polystyrene, polyester, or polycarbonate. The studies have also reported that the products of thermal degradation generally include the formation of H₂, CH₄, C₂H₆, C₂H_4, and C_9–46 HCs [15] and the pyrolysis oil obtained contains a range of hydrocarbons that can be used as a fuel source [[16], [17]].

Fig. 2.

(a) Parts of a commonly used surgical facial mask, (b) Parts of the 3-layered mask body.

The other possible conversion technologies to pyrolysis in handling biomedical waste are plasma pyrolysis and incineration. Plasma pyrolysis consumes greater than 2 KWh of energy to convert per kg of biomedical waste [18]. Incineration emits pollutants such as particulate matter, carbon monoxide, oxides of sulfur and nitrogen. These pollutants significantly impact the respiratory organ system of humans. Incinerators are polluting and plasma pyrolysis is energy intensive. Both of these technologies do not have any resource recovery possibility from biomedical plastic waste. Compared to both techniques, pyrolysis is energy efficient and operates at a significantly lower cost.

Thermal Degradation is a complex thermochemical process that is influenced by both physical and chemical factors [[19], [20],]. Additionally, carrying out the experimentation work is expensive and time-consuming; at times the scalability of these experiments is also very limited which makes the quantification process a cumbersome task. Additionally, it is a desire to obtain mathematical models and empirical relations to model the process, however, these equations from the first principle approach are often a computational task as it has stiff boundary conditions, which may vary depending upon the nature of the materials, operating conditions and other certain experimental rigidity which often limits the modeling process. However, the adaptation of an inter-disciplinary approach by using modeling techniques like Machine Learning (ML) has reduced these limitations and ameliorated the experimentation techniques adopted by the scientific community.

Machine Learning is a set of algorithms that are used to parse the data and learn without being explicitly programmed to use the data for prediction purposes. Depending upon the context of the application, ML has various algorithms namely Supervised, Semi-Supervised, Unsupervised, and Reinforced. Under the Supervised algorithm, the dataset includes labelled inputs and outputs and the algorithm identifies patterns in data, learns from observations, and makes predictions. These are used for Classification, Regression, and Forecasting type problems. Semi-supervised algorithms utilize both labelled and unlabeled data to learn about the unlabeled data. Unsupervised algorithms have parses on unlabeled data and attempt to organize the data in certain ways to identify correlations and relationships by analyzing the available data. The problems that have been frequently addressed fall under Clustering and Dimensionality Reduction. Reinforced Learning decides the best possible path to maximize the reward in a given situation based on the model's own experience. In this study, supervised ML is used as the data set obtained using TGA analysis is labelled. As cited in the literature, the ML approach to model the TGA process has been applied to model the thermal degradation process of fibres [13], biomass [[15], [21], [22]], crystals [23], and polymers [[24], [25], [26]].

In this study, the kinetic and thermodynamic study of the TGA process based upon data at a single heating rate value is undertaken. For the first time the application of an artificial neural network model to predict the thermal degradation behavior of (weight loss%) different components of surgical facial masks has been devised. The rational for implementing ANN based regression to predict the weight loss behavior of used face masks is originated from the lack of empirical and analytical models to predict similar results. In comparison with existing regression techniques ANN based method provides higher accuracy and more reliable. Additionally, this current study also reports thermal degradation kinetic parameters (Activation energy, pre-exponential factor), thermodynamic properties like enthalphy, entropy and gibbs free energy for all 3 components of used surgical masks (Ribbon, body, corner), which nowhere reported in the existing literature.

2.1. Materials and methodology

The methodology adopted in this study has been presented in Fig. 3 and the subsequent steps are explained in their respective sections.

Fig. 4.

Purelin transfer function.

Fig. 3.

Flow diagram of the adopted methodology.

2.2. Sample collection and experimental setup

The materials used to study the thermal degradation behavior were the different components of masks namely the mask ear ribbon, corners, and mask body. Biomedical waste plastics were collected from the Indian Institute of Technology (IIT) Madras and Srushti Fertility center & Women's Hospital, Chennai, Tamil Nadu, India. With utmost safety precautions, the collected biomedical wastes were then segregated into face masks, glucose bottles, IV tubes, syringes, and other PPEs such as head covers, aprons, etc. The medical wastes were exposed to UV radiation for disinfection on the hospital premises [18]. UV-disinfected wastes were preheated to 100–120 °C for secondary disinfection. The collected wastes were segregated and thermochemical conversion studies were conducted for evaluating the potential of resource recovery.

Facial masks intact were used for the analysis and the remaining was carefully discarded. The facial masks were disassembled into their individual components namely, the mask ribbon, corners, and mask body. The thermogravimetric (TG) analysis was carried out using TA Instruments SDT Q600, with temperatures ranging from 24 to 900°C under a nitrogen atmosphere with a purge gas flow of 100 mL.min⁻¹.

2.3. TGA analysis and thermal decomposition process

The thermal decomposition of solid materials is studied using thermogravimetric analysis, through which the variations obtained in conversion rate with time and temperature under different thermal conditions can provide degradation process information such as decomposition temperature, peak temperature, peak, and average reaction rate. The thermal degradation of solids under non-isothermal can be mathematically represented by,

$$\frac{d\alpha }{dT}=\frac{A}{\beta }{e}^{-\frac{E}{RT}}f\left(\alpha \right)$$ (1)

$$g\left(\alpha \right)=\frac{A}{\beta }\underset{0}{\overset{T}{\int }}{e}^{-\frac{E}{RT}}dT$$ (2)

Replacing the $\left[-\frac{E}{RT}\right]$ as x and transforming the limits in Eq. (2), it becomes

$$g\left(\alpha \right)=\frac{AE}{\beta RT}\underset{x}{\overset{\infty }{\int }}\frac{e^{-x}}{x^2}dx$$ (3)

The exponential integral is replaced with p(x), which has no analytical solution and is expressed by approximations.

$$g\left(\alpha \right)=\frac{AE}{\beta RT}p\left(x\right)$$ (4)

Where, f(α) is the reaction model in its differential form, g(α) is the reaction model integrated form, β is the heating rate (in K/min), A is the pre-exponential factor, T is the temperature (in Kelvin), t is the time, E is the activation energy (J.mol ⁻¹) and R is the universal gas constant (R = 8.314 J. K⁻¹. mol⁻¹) and α is the fractional weight loss, which is given by, $\alpha =\frac{m_0-m}{m_0-m_{\infty }}$ and m ₀, m _∞, m, represents the initial weight, final weight (after TGA analysis), and current weight (at the time instance during TGA analysis). The values of kinetic parameters such as activation energy and pre-exponential factor can be determined using the conversion data by deploying model-free methods such as Kissinger-Akahira-Sunose (KAS), Flynn-Wall-Ozawa (FWO), and Advanced Vyazovkin method or by using model-fitting methods such as Coats-Redfern (CR) method, which is used in this study.

2.4. Coats-Redfern method

In this study, the Coats-Redfern method was used to model the thermogravimetric process. This is an integral model-fitting method [[27], [28], [29]] that involves fitting various reaction models to the extent of the reaction-temperature curves and simultaneously determining the activation energy and frequency factor. This method uses the asymptotic series expansion for approximating the exponential integral p(x) in Eq. (4) and hence becomes,

$$\ln \left(\frac{g\left(\alpha \right)}{T^2}\right)=\ln \left[\frac{AR}{\beta E}\left(1-\frac{2R\overline{T}}{E}\right)\right]-\frac{E}{RT}$$ (5)

Where, $\overline{T}$ is the mean experimental temperature. Plotting $\ln \left[\frac{g\left(\alpha \right)}{T^2}\right]$ vs.$\left[\frac{1}{T}\right]$ of Eq. (5), gives E and A from the slope and intercept, respectively. Various reaction model approximations are available in the literature for g(α), for the purpose of this study authors have taken the sigmoidal rate kinetic model wherein the rate determining mechanism assumes random nucleation and its subsequent growth. Under this g(α) is approximated by,

$$g\left(\alpha \right)=-\ln \left(1-\alpha \right)$$ (6)

By substituting Eq. (6) in Eq. (5), we obtain the following Eq. (7).

$$\ln \left(-\frac{\ln \left(1-\alpha \right)}{{\mathrm{T}}^2}\right)=\ln \left[\frac{\text{AR}}{\beta \mathrm{E}}\left(1-\frac{2\mathrm{R}\overline{\mathrm{T}}}{\mathrm{E}}\right)\right]-\frac{\mathrm{E}}{\text{RT}}$$ (7)

The above-mentioned equations are attempted to fit a straight line of the form,

$$y=c+mx$$ (8)

Where the left-hand side denotes points on y, whereas$\left(-\frac{1}{T}\right)$, $\ln \left[\frac{AR}{\beta E}\left(1-\frac{2R\overline{T}}{E}\right)\right]$, and $\frac{E_a}{R}$ represent (x), intercept (c) and slope (m), respectively. Once obtaining the best regression fit, thereafter the values of the activation energy (E_a) and pre-exponential factor (A) is calculated from the slope and intercept values.

2.5. Thermodynamic consideration of the thermal degradation process

The thermodynamic analysis of any chemical process provides a better understanding of the overall performance of the process under study, through which eventually sources of losses due to irreversibility could be identified. In our study, we attempt to perform the thermodynamic analysis of the thermal degradation process of different sections of the mask, by using the determined kinetic parameters to calculate the values of thermodynamic parameters such as the changes in Enthalpy (ΔH, kJ/mol), Entropy (ΔS, kJ/mol.K) and Gibbs Free Energy (ΔG,  kJ/mol). Additionally, the thermodynamic parameters are crucial to perform measurements of energy and evaluating the direction and feasibility of the thermal degradation process [30]. The values of ΔH, ΔG and ΔS can be calculated using the following equations,

$$\Delta H=E-RT$$ (9)

$$\Delta G=E+RT\ln \left(\frac{kT}{hA}\right)$$ (10)

$$\Delta S=\frac{\Delta H-\Delta G}{T}$$ (11)

Where T represents the temperature (K), k_B is the Boltzmann constant (1.381 ×  10⁻²³  J/K), h is the Planck constant (6.626 ×  10⁻³⁴  J.s) and A is the pre-exponential factor, respectively.

2.6. Machine learning approach to model the thermal degradation of mask

Machine learning models are widely used for complex regression and prediction tasks owing to the pattern learning tendencies of the algorithms, which otherwise would have been a computationally intensive task, especially in studies where a lot of interdependencies amongst various variables and process parameters are involved such as in this study. The study presented in this work belongs to the supervised class of machine learning as a labelled data set obtained from the TGA instrument is used. Under the supervised class, various algorithms such as k-Nearest Neighbor, Naive Bayes, Decision Trees, Linear Regression, Support Vector Machines (SVM), and Neural Networks can be used for the regression analysis. Amongst all, for the purpose of this study Artificial Neural Network (ANN) is used as they can model data with high volatility and non-constant variance, given its ability to learn hidden relationships in the data without imposing any fixed relationships in the data.

Typically, an ANN is inspired by the neuron connections in the brain that are capable of learning and capturing patterns for predictions using empirical models. ANN is conventionally constructed in a three-layered structure, i.e., the input, hidden layers, and the output. Each layer has a different number of computational elements called neurons. The optimal number of neurons in the hidden layers depends on the nature of the problem and the training algorithm used. Prior to the modeling process, it is necessary to review the quality of the dataset, the principal parameters, the training algorithm, and the number of neurons in the hidden layer. In this study, a conventional feed-forward backpropagation ANN was deployed as it had the lowest error compared with other network architectures (cascade, Elman, competitive, etc.). The algorithms namely Levenberg-Marquardt, Bayesian regularization, and scaled conjugate gradient are most commonly used for training the network. In this study, Levenberg-Marquardt was chosen for this work because of the high performance in the data processing. The output of any neural connection is obtained by passing the data points from transfer functions, which renders the input signals to output signals, that are used to describe the curve function of the experimental data. Various transfer functions are as hyperbolic tangent sigmoid[31], purelin, and logarithm sigmoid, whose logical relation is as follows:

$$Y_j=f_t\left(\sum _{i=1}^n{w}_{ji}{x}_i+b_j\right)$$ (12)

Where Y_j is the output of the jth neuron, w_ji is the weight of the jth neuron in the ith input, x_i is the independent variable, b_j is the jth neuron bias, n is the number of the input variable, and f_t is the transfer function. In this study, purelin was used as the transfer function used that considers and describes the complex relationship between the polymers in the different components of the mask, time, temperature, and heating rate to the mass loss behavior.

The time (t), temperature (T), and heat flow (Q) were taken as the input data and the Mass Loss (%) was considered as the output. Input data points to the ANN are allotted into three data sets including training, validation, and test. Table 1 report the data used in the ANN development.

Table 1.

The range of data employed in ANN training.

Parameter	Type	Minimum value			Maximum value
		Ribbon	Body	Corner	Ribbon	Body	Corner
Temperature (K)	Input	297.15	298.12	298.70	946.19	740.1	871.15
Time (min)	Input	0.70	0.5	0.7	65	45	57
Heat flow (mW)	Input	0.01	0.06	0.037	73.59	43	40
Mass Loss (%)	Output	0.77	0	0	100	100	100

Several numbers of hidden layers and training iterations are performed to obtain the optimal number of hidden layers and result in a low error. Obtaining the optimality is necessary here as a lower number of hidden layers represent simple relations in the data whereas a large number of hidden layers can learn complex data patterns. Increasing the number of hidden layers may lead to overfitting of the data, which generally happens when a model learns too many details along with the noise from the training data, due to this the network fails to generalize the patterns present in the training dataset. In the data pre-processing stage, the redundant values and inappropriate parameter measurements were filtered to improve the quality of the dataset. In our present study, we used MATLAB R2021b to model the Artificial Neural Network and Microsoft Excel 2019 for data preparation and pre-processing.

The performance of the model was characterized using the coefficient of determination (R ²), which is the proportion of the variance in the output variable that is predicted from the independent variable.

$$R^2=1-\frac{{\sum }_{i=1}^n{\left(x^{\prime }\left(i\right)-x\left(i\right)\right)}^2}{{\sum }_{i=1}^n{\left(x\left(i\right)-\overline{x}\left(i\right)\right)}^2}$$ (13)

Where x′(i)the predicted mass loss (%) is, x(i) represents the experimental value, $\overline{x}$ is the mean value, n is the number of test data and i is the individual values.

3. Results and discussions

The components of surgical masks comprise various polymers which have different decomposition temperatures and reactivity toward heat. The Derivative Weight (%/degK) basically denotes the change in mass with change in temperature. Fig. 6 shows the Derivative. Weight (%/degK) and reaction rate as a function of temperature for the different components of the masks. More detailed information about the peak temperature and their respective conversion is represented in Table 2 .

Fig. 6.

(a) Variation of Derived Weight (%/ °C) and (b) Conversion α with temperature.

Table 2.

The values of peak temperature with respective conversion and weight%.

Mask Component	Peak Label	Peak Temperature (TP, K)	Conversion at TP	Weight (%)
Ribbon	2	431.62	0.0028	99.75
	3	649.57	0.1121	88.92
	4	718.96	0.6575	33.02
Body	1	428.10	0.0125	98.70
	5	727.42	0.6896	29.80
Corner	6	731.47	0.5408	42.72

For all the three mask components, the thermal decomposition plots followed a similar trend [29], in which the initial weight loss is due to the loss of intrinsic moisture and high volatility plasticizers from the polymer. From Fig. 7, it is observed that in the body of the mask for a temperature around 600 K, the mass has reduced to nearly 96% whereas at the same temperature the mass loss is significantly less in the ribbon and corner section, the mass is reduced to 98.86% and 99.08%, respectively. In the second phase of the decomposition process, weight loss occurs due to the degradation of the residual polymer. In general, the mask ribbon mainly consists of polyurethane and nylon (primarily comprising of polyamide), whose decomposition temperature lies in the temperature range of 473–843 (Peak 2) and 473–753 K (Peak 3 and Peak 4), respectively. Maximum mass was lost up to 33% till the peak temperature of 718.96 K was reached. In the body section, two peaks Peak 1 and Peak 5 occur at 428.10 K and 727.42 K, respectively, at these temperatures the body observed a mass reduction from 98.70% up to 29.80%. In the thermal degradation of the corner section, a sharp peak at a temperature of 731.47 K was obtained up to which the mass was reduced to nearly 43%. During the thermal degradation of the polymers in the mask components, weight loss may also occur due to the combustion of compounds such as residual carbon from polymer breakdown, leaving behind the ash residue due to the inorganic filler materials present in the respective components.

Fig. 7.

Variation Mass Loss (%) of mask component with temperature.

3.1. Estimation of kinetic parameters

The quantification of reaction rate kinetics signals toward the determination of reaction order, activation energy, and pre-exponential factor. The kinetic parameters were estimated using the Coats-Redfern method[27], which typically involves a straight line curve fitting procedure. The pre-exponential factor (A) which is an indication of the frequency of collisions between reactant molecules, depicted an increasing trend with temperature for all the mask components, which can be addressed by the fact that increasing temperatures result in an increment in kinetic energy of the molecules. The estimation of parameters such as Activation Energy (E) and pre-exponential factor (A) from the line of best fit (as shown in Fig. 8) are reported in Table 3 . The coefficient of determination (R ²) > 0.97 for the lines of best fit for all the components. The average values of activation energies obtained using the slope of the best fit line for the Ribbon, body, and corners were 35.544 kJ/mol, 23.9 kJ/mol, and 27.3 kJ/mol, respectively.

Fig. 8.

Line of best fit for the determination of Kinetic parameters.

Table 3.

Estimation of kinetic parameters.

Mask component	Peak Temperature (TP, K)	α at TP	Slope	Intercept	E (kJ. mol−1)	Pre-exponential factor (A × 107)
Ribbon	718.96	0.6575	4275.1	9.40	40.185	6.668
Body	727.42	0.6896	2866.2	13.76	23.829	2.967
Corner	731.47	0.5408	3290.8	13.90	27.259	1.338

3.2. Thermodynamic analysis of thermal degradation of mask components

The change in enthalpy (ΔH), Gibbs free energy(ΔG), entropy (ΔS), and for the different mask components as a function of the conversion rate α is shown in Fig. 9. It is observed that the values of the thermodynamic parameters nearly follow a similar trend, where ΔH and ΔG decrease smoothly as the conversion rate increases. ΔS Increases with conversion rate, but overall, it remains nearly constant throughout the entire thermal degradation process. The average values of the thermodynamic parameters of the mask components have been summarized in Table 4 .

Table 5.

ANN training parameters.

Parameters	Value
	Ribbon	Body	Corner
Network type	Feed-forward backpropagation
Training algorithm	Levenberg-Marquardt
Number of input nodes	3	3	3
Number of hidden layers	5	7	7
Number of output nodes	1	1	1
Number of epochs	82	9	14
Validation checks	49	6	6
Mu	1.00e-07	1.00e-04	1.00e-05

Fig. 9.

Plots of ΔH, ΔG and ΔS with respect to the conversion rate (α) for (a) Ribbon, (b) Body, (c) Corner.

Table 4.

Average values of different thermodynamic parameters for the mask component.

Mask Component	ΔHavg (kJ.mol−1)	ΔGavg (kJ.mol−1)	ΔSavg (J.mol−1.K−1)
Ribbon	349.67	411.83	-90.17
Body	187.58	224.27	-54.39
Corner	22.65	70.20	-65.41

The change in enthalpy (ΔH) correlates to the energy difference between the reactants and the activated complex and can be used to assess the favourability of the thermal degradation process [32]. For the entire mask components, the positive value of ΔH signifies the endothermic nature of the thermal degradation process in an inert environment. From Fig. 9 it is evident for all the mask components the ΔH decreases with the increasing reaction conversion rate. ΔG represents the total energy rise of the reaction system during the thermal degradation process and can be used to evaluate the direction and reaction feasibility [33]. Higher values of ΔG corresponds to the lower probabilities of the occurrence of the reactions and lower values signify the likelihood of the reactions occurring more easily. With ΔG values, the spontaneity of the process can be determined; processes with positive ΔG values mean in the process non-spontaneous reactions are involved. As evident from Fig. 9, it is clear that for all the mask components with increasing α the ΔG depicts the increasing trend which implies the reduced occurrence of the reaction. The change in entropy ΔS of a process can be used to estimate the degree of disorder of a system, in this study the thermal degradation [34]. Lower values of ΔS imply that the process can attain thermodynamic equilibrium and lower reactivity. For all the components the ΔS values decrease with conversion rate, signifying the lower reactivity of reactants at higher temperatures and higher reaction conversions. However, ΔS remains fairly constant throughout the thermal process.

3.3. Machine learning model for the thermal degradation process

In order to obtain an accurate predictive model using a data-driven technique such as Neural Network, it is essential to identify the crucial parameters controlling the process model and understand their dependency on one another, one way to proceed is using heat maps. A heat map is a two-dimensional data visualization tool that essentially shows the degree of correlation among two variables pertaining to a dataset. Patterns among different variables can be identified by observing how individual cell colors change in the plot [35]. Fig. 10 shows our attempt to use a heat map for our work to visualize the correlation pattern and dependency among the parameters pairwise. The heat maps in Fig. 10 for the different components show a color scaling across parameters; this signifies the similarity in the thermal degradation process for components. The parameters that are considered for the present study are time (t), temperature (T), and heat flow (Q), and their effect on mass loss (%) is studied.

Fig. 10.

Heat map depicting correlation pattern among the parameters in (a) Ribbon, (b) Body, and (c) Corners.

Each squares in the heat map shows a correlation between the parameters, where the correlation ranges from -1 to +1. A correlation close to 1 indicates strong positive correlation amongst the parameters that is as one increases so does the other. Values closer to zero means there is no linear trend between the two parameters. A correlation closer to “-1″ signifies that by increasing one parameter will decrease as the other increases. The plot is symmetrical about the diagonal since the same two variables are being paired together in those squares. The diagonals all have a correlation coefficient of 1 because those squares are correlating each variable to itself hence it's a perfect correlation. Considering the similarity of the correlation patterns amongst the Ribbon, Body and Corner, the interpretations of all are the same. To re-iterate, the variable shown in the heat maps in Fig. 10, are t, T, Q ^. , Mass (%) is the time, temperature, heat flow and mass loss%, respectively. The temperature shows a strongly linear variation with time, which is self-explanatory as in the heating process; the sample is provided heat at a fixed heating rate. With both time and temperature, the mass (%) shows a negative correlation pattern as during the heating process with the passage of time, the temperature increases which results in the components of the undergoing chemical reactions and thereby decreasing in available mass of the sample

Pertaining to three different mask components, individual ANN models for each was implemented using MATLAB R2021b, based on the architecture shown in Fig. 5 . The trial and error approach method was used to find the optimal number of hidden neurons. An equally sized dataset of 700 data points was used to train the three ANN models using the Levenberg-Marquardt algorithm with an epoch cycle of 600. The dataset was divided in a Train-Test-Validation split at a 70-20-10 ratio, i.e., 490 data points were used for training the network, 140 data points were used in testing and the remaining 70 unseen data points were used to test the accuracy of the trained network.

Fig. 5.

Schematic of a single hidden layer, three inputs, and one output-based feedforward ANN architecture used in this study.

In the training of the neural network, an epoch refers to one cycle of passing an entire dataset through the training process back and forth the network. However, an optimal epoch is to be determined as with increasing the epochs, the training data is fed to the neural network in different patterns enabling a better generalization and pattern recognition within the data whereas on the contrary lesser epoch leads to under fitting. As the number of epochs increases, a greater number of times the weights of the neural connection are changed in the neural network and the curve go from under fitting to optimal to overfitting curve. In the MATLAB environment, the epochs are self-adjusted during the training process until the required convergence criteria are met. In order to prevent overfitting during the training, validation checks are performed that provide an unbiased evaluation of a model fit, based on which the weights are adjusted. Mu is referred to as the momentum that shows the way to adjust the learning rate. During the training process, the choice of weights and biases is optimized by minimizing the sum of the squared errors by using the method of steepest descent. However, it is desired to slow the speed of the descent so that the search value does not fly back and forth across the minimum without stopping sufficiently near it, hence the Mu value aids in that. MATLAB Neural Network toolbox allows specifying the range of Mu value or uses a default range to train the network and the former was used to train the network in this study.

During the training process, the conditions of hyper parameters are evaluated to prevent overfitting, one such hyper parameter is the number of hidden layers. In Table 6 , the variation of R ² with the number of hidden layers is observed and if the R ² value is 1, then that particular hidden layer number is discarded as it corresponds to the over-fitting of the network. Hence, the optimal numbers of hidden layers are 5, 7, and 7 for neural network models corresponding to Ribbon, Body, and Corner, respectively. Corresponding to these optimal hidden layers, along with the R ² value, other performance metrics such as MSE and RMSE have also been provided in Table 7 . The R ² ranges from 0 to 1, and the MSE, RMSE ranges from 0 to infinity, however as a rule of thumb, when the R ²-value approaches 1 or the model attains a perfect fitting, the values of MSE and RMSE approaches 0.

Table 6.

Variation of R² with number of hidden layers keeping other parameters constant.

	Number of Hidden Layers
Component	3	4	5	6	7	8	9
Ribbon	0.9839	0.9963	0.9989	1.0000	1.0000	1.0000	1.0000
Body	0.9834	0.9886	0.9917	0.9932	0.9968	1.0000	1.0000
Corner	0.9798	0.9829	0.9861	0.9965	0.9985	1.0000	1.0000

Table 7.

Performance Metrics of the ANN model of the different components of the mask.

Component	Ribbon	Body	Corner
Performance Metric
R2	0.9989	0.9968	0.9985
MSE	1.1044	1.0353	1.0986
RMSE	1.0508	1.0175	1.0481

Fig. 11 shows the ANN regression plots for each of the mask components. The regression ANN predicts the output predicted values as a function of the experimental datasets. Besides, the linear function of the prediction output versus experimental data confirms the good fit with a high coefficient of determination (R ² > 0.99).

Fig. 11.

Regression plots between experimental and predictive results, (a) Ribbon (b) Body (c) Corner.

4. Conclusion

Properly disinfecting and segregating used surgical masks based on their individual components like Body, Ribbon, and Thread is critical to effective waste utilization and catalytic pyrolysis and other approaches. The composition of individual components place a critical role in determining operating temperature and pressure of pyrolysis reactor and also avoid contamination of derived product due to different composition of the feed stock. Selection of ideal operating conditions also helps to minimize the cost of waste management and products recovery. Since the fight against the pandemic is still undergoing, the effective management of COVID-19 waste disposal and resource recovery is of paramount importance.

In the present study, thermal degradation of used surgical mask components- ribbon, body, and corner was undertaken by conducting Thermogravimetric Analysis. Using the TG data, kinetic parameters like activation energy and pre-exponential factor were estimated using the Coats-Redfern method. Thermodynamic analysis of the mask components was also undertaken and values of ΔH, ΔG, and ΔS were found. Machine learning approach based Artificial Neural Network (ANN) model was built for each component to predict the TG behavior, respectively. The predicted data obtained using ANN was at high accuracy, as each of the models had, R ² > 0.99. Upon establishing an optimized model, any input data can be easily included in trained ANN and the respective output can be appropriately obtained.

Funding

This research article did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.