1. Introduction
Researchers are currently shifting towards a powerful artificial neural network (ANN) model analysis to handle linear and non-linear systems. Therefore, in this work, ANN will be implemented as new technique to analyze the pyrolysis reaction of PP. Muravyev et al. (2021) [
1] reported that there is no mathematical tool that can solve all the pyrolysis TGA problems. In some cases, the ANN may become an alternative option.
There are two earlier-reported publications, including my own publication, for TGA data analysis using ANN [
2,
3,
4]. In this research, Dubdub (2022) [
4] summarized and presented most of the reported publication list from before the year of 2022 with different architecture details for non-isothermal TGA data.
In this introduction, all the papers that used the ANN technique for TGA data and were published within the current year (2022) have been reviewed. Usually, ANN networks agree to be encoded numerically, as “(x-y-z).” For the neurons, these x, y, and z variables denote the input layer, hidden layer, and output layer (Muravyev et al. (2021) [
1]), respectively. Sometimes, more than three numbers exist, such as “(x-y-z-a),” which means there are two hidden layers with y, z neurons. This information is mentioned in the early stages of this paper because some of these codes will be used in the introduction section.
Khan and Taqvi (2022) [
5] reviewed the applicability and various benefits of machine learning methods such as in prediction, fault detection, optimization, and quality control. They classified ML into two main types: supervised learning techniques, including ANN, and unsupervised learning techniques. The authors also mentioned different applications of ANN in Chemical Engineering process systems, including chemical reaction processes, the auto tuning of PI and PID controllers, coagulation in poly aluminum chloride, water-treatment plants, the optimization and modeling of cross-flow ultrafiltration, the performance of wastewater treatment processes, oil degradation in wastewater, water quality in terms of chemical oxygen demand (COD) and biological oxygen demand BOD [
6], and the removal of zinc ions from wastewater in hydrodynamic cavitation for biomass pretreatment.
Gonzalez-Aguilar et al. (2022) [
7] reported the effects of heating rate and temperature on the thermal pyrolysis of expanded polystyrene with respect to liquid conversion yields and physicochemical properties in a semi-batch reactor using factorial statistical analysis.
Demir (2022) [
8] used different compositions of polycaprolactone (PCL)/polyvinyl chloride (PVC) blends in order to study the thermal degradation kinetics with a four-heating rate using ANN analysis. Further, the investigated results of the decomposition temperature, heating rate, and percentage of PVC in blends as input data for the ANN model, and the percentage of weight remaining during degradation as output data for the ANN model were reported by [
8]. He concluded the 3-10-10-1 network with Logsig–Tansig transfer function with feed-forward backpropagation was the best network for this system. He then developed another system with a 4-10-10-1 topology with a Logsig–Logsig transfer function by taking the degradation temperature, heating rate, percentage PVC, and percentage weight left data as input data, and using the activation energy values as output data.
Ai et al. (2022) [
9] investigated the use of ANN for the co-pyrolysis of oily sludge and high-density polyethylene (HDPE). They used two models: one that predicted the interactive effect of HDPE using the heating rate, temperature, and mixing ratio as the input variables, and a second model that predicted the activation energy with the mixing ratio and reaction conversion degree as the input variables. The performance and validation of the model was determined by the average regression coefficient (R
2) and by measuring the root mean squared error (RMSE) calculation. For both cases, authors used two hidden layers with (8–10) neurons (R
2 = 0.99) and (10–7) neurons (R
2 = 0.92) for the first and second models, respectively.
Balsora et al. (2022) [
10] studied the pyrolysis of lignocellulosic biomass by an ANN. They used the preliminary analyses, including proximate, ultimate, and biochemical analysis, as input variables (four different input set variables) and the kinetic parameters as output variables by four different models of an ANN network. They used the mean square error (MSE) statistical-criteria parameter to obtain the optimum number of neurons in the hidden layer and found 9, 6, 200, and 39 neurons for ANN-1, ANN-2, ANN
-3, and ANN-4, respectively. They tried to improve the performance of the first two ANN models by combing the number of input variables to obtain the third ANN model, and they improved further in the fourth ANN model by including the biochemical analysis as additional input for the input variables of the third ANN model. As a sensitivity analysis study, they highlighted that the effect of the biomass analysis on the output variables estimation was very significant. Jacob et al. (2022) [
11] studied the ANN to predict the mass loss of a raw-mustard biomass using three hidden layers and one output layer. They used the mixing ratio, heating rate, and temperature as input variables and the mass loss as only output variable.
Kartal and Özveren et al. (2022) [
12] applied the ANN to the pyrolysis of almond shell and imbat coal. They focused their work on estimating the activation energy for the two samples by using the ultimate analysis and the experimental TGA data for the ANN model. These inputs include the carbon content, oxygen content, hydrogen content, temperature onset, temperature end, heating rate, flowrate of gas, type of thermochemical process, and particle size ranges. They used, out of total 188 datasets, 80% data for training, 5% for validation, and the remaining 15% for testing. Furthermore, they used a Tansig activation function for the two hidden layers with (12-6) neurons in the first and second hidden layers, respectively.
Khodaparasti et al. (2022) [
13] used (2-8-1) as network architecture for an ANN network, with the temperature and heating rate as input variables and the remaining mass (%) as the target or output variable for the co-pyrolysis of municipal sewage sludge and microalgae chlorella vulgaris. They found that the R
2 value for this network was close to one, indicating that the ANN prediction was very close to the experiential data.
Li et al. (2022) [
14] applied the ANN (multi-layer, feed-forward, network-backpropagation) for one run with a heating rate 20 K min
−1 for the co-pyrolysis of agricultural waste and HDPE. They applied the mass change of feedstock as the output variable and temperature as the input variable. They concluded the best architecture was three neurons in one hidden layer with a sigmoid function.
Nawaz and Kumar et al. (2022) [
15] studied the ANN for the pyrolysis of a biomass (Sesbania bispinosa). They found that the best structure was (2-10-1) by considering R
2 and MSE as the main criteria. Similar to many other papers, they used the temperature and heating rate as input variables and the weight-loss percentage as the output variable.
Postawa et al. (2022) [
16] used the ANN differently when compared to the above-mentioned researchers by selecting four parameters: the heating rate
, activation energy, pre-exponential factor, and reaction order as input variables and choosing the percentage shares of cellulose, hemicellulose, and lignin as output variables. The result was a (4-11-16-3) structure in which two hidden layers of 11 and 16 neurons were considered to be the first and second hidden layers, respectively.
Dubdub and Al-Yaari (2022) [
4] and Dubdub and Al-Yaari (2021) [
17] employed the same ANN technique to check the validity for TGA data. They used a feed-forward backpropagation ANN model with two hidden layers to predict the TGA data. In the first cited paper [
4], they applied two input variables: the heating rate (K min
−1) and the temperature (K); in the second paper, the ratio between the catalyst and the polymer was included as a third input [
17].
In this present paper, a highly efficient, developed ANN model has portended the pyrolysis behavior of PP polymer using TGA data. Moreover, a sensitivity analysis has been conducted to find the uncertainty in the relationship between the input variables and the output variables. Finally, temperature has been found to be a more sensitive input parameter when compared to the heating rate variable.