The 157 tobacco samples originated from Brazil, Zimbabwe, and 104 counties in six provinces in China’s major tobacco producing regions, including Henan, Yunnan, and Guizhou provinces. The tobacco collection years included 2017, 2018, and 2019. Tobacco can be classified into 35 grades according to the national standard GB 2635-1992.After being placed in a constant temperature and humidity chamber with a temperature of (22 ± 1) °C and relative humidity of (60 ± 2)% for 48 h to reach equilibrium, the tobacco leaf samples were pulverized by means of a high-speed grinder and screened by a 60 mesh (250 μm) sieve for further use.

A total of 157 tobacco samples were studied. The chemical information included total phytoaloids, reducing sugars, total sugars, total nitrogen, potassium, chloride, starch, dichloromethane extract, solanasol, phosphate, magnesium, calcium, polyphenols, refractory acid, total amino acids, amadori compounds, neophytadiene, and PH. Tobacco samples were treated as solution according to tobacco industry standards and then analysed directly for total phytoaloids, reducing sugars, total sugars, total nitrogen, potassium, chloride, starch, dichloromethane extract, phosphate, magnesium, calcium, refractory acids, total amino acids, amadori compounds, neophytadiene, using a flow analyser (Alliance-Futura). The content of polyphenols, solanasol were determined using a liquid chromatograph. PH values were measured by a Mettler-Toledo Seven Compact PH meter.

In the analysis of pyrolysis characteristic parameters of tobacco leaf samples, it is necessary to confirm the reliability of the pyrolysis characteristic data, which relies on a good repeatability of thermogravimetric experiments. Figure 1 illustrates the DTG curves of three repeated experiments carried out on a particular sample, and it can be visually observed that there are no significant differences between the three experiments. In addition, the differences between the three experiments are described quantitatively using the Normalized Root Mean Squared Error (NRMSE). NRMSE between trial 1 and trial 2 is 1.85%, and NRMSE between trial 1 and trial 3 is 0.57%. It indicates that there is a good repeatability for the thermogravimetric experiments, which is sufficient to meet the experimental requirements.

Since the temperature intervals of the original differential thermogravimetric data (DTG) are not equal, interpolation is required to obtain thermogravimetric data from different tobacco samples at the same temperature points. The temperature range of the interpolation is from 105°C to 900°C, with a temperature interval of 0.1°C. The DTG curves of 157 samples are shown in Figure 2, with a total of 8436 points obtained for each sample. As can be seen from Figure 2, in general, the DTG curves of the 157 tobacco samples were similar in shape. The differences in the DTG curves between the tobacco samples were mainly reflected before 500°C, with less variability in the rest of the temperature range. This is because there are two distinct stages in the pyrolysis process of tobacco. The first stage is 100°C–230°C, this stage is mainly monosaccharides, free amino acids and other thermally unstable, volatile components degradation. The second stage is at 230°C–500°C, which is mainly the pyrolysis of biological components such as hemicellulose, cellulose and lignin of tobacco species (Barontini et al., 2013; Ma et al., 2022).

For each chemical index, the maximum and minimum values were counted, then the distribution range was divided into a number of intervals, the number of samples in each interval was counted, and the distribution probability was calculated to make a densities plot of the distribution of the chemical content of tobacco samples. Figure 3 shows the distribution densities of 19 chemical constituents of 157 tobacco samples. The horizontal axis represents the interval of content values, and the vertical axis represents the probability density distribution. It can be seen from the figure that the distributions are not uniform for various chemical constituents.

The linear correlation coefficients among 19 chemical components of 157 tobacco samples are shown in Figure 4. The values in the heat map reflect the strength of the correlation between the chemical components (the higher the value, the stronger the correlation). It can be seen that the correlation coefficients between most components are very low, and the correlation coefficients above 0.9 are only total sugar, reducing sugar, methylene chloride extract, solanesol, calcium and nonvolatile acid. These results suggest that each chemical components needs to be modelled independently due to the dependence between different chemical components is minimal. In the end, we considered all 19 chemical information as features as inputs for machine learning.

To explore quantitative analysis models between DTG data and chemical components, Partial Least Squares (PLS), Support Vector Regression (SVR), Gaussian Process Regression (GPR), Multiple Linear Regression (MLS), Random Forest (RF) and Shallow Neural Network (SNN) were used to establish fitting models for the representative chemical components of tobacco materials, including total alkaloids, reducing sugars, and total nitrogen. The regression performance of each model on the training set and validation set are shown in Table 1. It can be observed that for the three chemical components, the R ² of the MPL, RF and RNN model training sets and test sets are below 0.7, indicating poor fitting performance for these three models on the high-dimensional DTG data. Among them, the R ² of MPL for the three chemical components are all less than zero, indicating that linear regression cannot fit DTG data effectively. On the other hand, the R2 of the training set and test set of the PLS, SVR and GPR models are all above 0.7, indicating good fitting and prediction performance relatively. Therefore, PLS, SVR and GPR models are selected to build quantitative analysis models between DTG data and 19 chemical components for comparison in the following analysis.

It is crucial to select the number of latent variables in the PLS modeling process. The model will be overfitted when there are too many latent variables, while selecting too few latent variables will result in loss of sample information and insufficient model fitting. Therefore, the number of latent variables that cumulative contribution rate of each variable reached 90% was set in the experiment, and considering the root mean square error of cross-validation (RMSECV), the number of latent variables with a cumulative contribution rate of about 90% and relatively small RMSECV were selected comprehensively.

When establishing the SVR model, the cross-validation loss is used as the goal to find the optimal penalty parameters. Considering that the GPR model performance of different kernel functions may be different, the square exponential kernel (SE), exponential kernel (Exp) and rational quadratic kernel (RQ) were selected to build models respectively, and the performance average of the three models were taken as the output result of the GPR model.

The performance comparison of the PLS, SVR and GPR models is shown in Table 2. It can be observed that for the 19 components predicted, the PLS model has higher R ² values (0.76–0.99) and most of them are above 0.85. Moreover, it has lower RMSE for most chemical components, indicating its optimal performance. For 19 chemical components, the performance of SVR and GPR models are similar, with only a few R ² values above 0.8 and most R ² values below 0.75, but some RMSE values are smaller than PLS. The R ² variation between the training sets and test sets of PLS is small, indicating that the model effectively balances fitting performance and generalization ability when selecting the number of latent variables.

Figure 5 shows the fitting performance of the three models for each chemical composition in the training and test sets (only three compositions are listed, the details of the models for the other chemical compositions are presented in Table 2). Where the x-axis is the true value, the y-axis is the model prediction, and the black diagonal line is the best fit line. It can be seen that of the three constituent predictions listed, compared to the other two models, the training and test sets of the PLS model have a relatively high and high agreement between the prediction results and the ground truth, which is closer to the black diagonal line. This indicates excellent fitting and prediction performance with high accuracy. In addition, it can be seen that under the three models, the difference in R ² and RMSE between the test set and the training set is not large, indicating that the models are not overfitted. For the PLS model in the vast majority of the above discussion shows that the PLS model used can effectively characterize the complex relationship between DTG curves and the chemical composition of tobacco materials. The PLS model with higher accuracy can be further used to investigate the relationship between the content of other components and the pyrolysis process.

PLS is a supervised multivariate statistical analysis method that combines the variability information of the dependent variable to extract latent variables. It can describe the degree and direction of the dependence of each latent variable on the original variables. The VIP (Variable importance in the projection) scores can be obtained to evaluate the contribution and explanatory ability of each variable to the model by calculating the variable coefficients, model weights and residuals of PLS regression. A higher VIP score indicates a more important variable that has a greater impact on the overall model. Given the good performance of the PLS regression model, calculating the VIP scores of the PLS regression model for the DTG of tobacco materials and the content of chemical components can help identify and select temperature points that contribute to the prediction of chemical component content significantly. This allows for the identification of temperature ranges that influence the chemical component content during pyrolysis. Generally, a VIP score greater than 1 indicates that it has important influence on the model.

The VIP score plot of the temperature points for chemical components is shown in Figure 6, which shows the VIP score plots for total alkaloids, reducing sugars and total nitrogen. The x-axis represents the temperature points, starting from 105°C. From the figure, it can be observed that the VIP scores of the temperature points for total alkaloids greater than 1 appeared within the range of 135°C–263°C and 332°C–385°C, and four distinct peaks appeared in the range of 135°C–263°C. The VIP scores of the temperature points for total nitrogen greater than 1 appeared within the range of 153°C–246°C and 260°C–399°C. The VIP scores of the temperature points for reducing sugars greater than 1 appeared within the range of 150°C–390°C, 510°C–520°C, and 688°C–701°C, and the peaks occurred near the temperature points of 135°C, 215°C, 345°C, respectively. It indicates that the pyrolysis rate in these temperature ranges has a significant impact on the regression model for the corresponding component content, especially near the peaks.

The pyrolysis temperature ranges of VIP scores greater than 1 for the 19 chemical components are shown in Table 3. It indicates that different temperature ranges have significant influences on the regression of different components of tobacco materials. The temperature ranges of VIP scores greater than 1 for chemical components are mostly within the range of 130°C–400°C, and a few of them are above 400°C. Moreover, the temperature ranges above 400°C are relatively short, indicating that the temperature ranges that have the most impact on the regression of chemical components of tobacco materials are mostly below 400°C. Combined with the samples DTG curve in Figure 2, it can be seen that the differences in the DTG curves between samples are mostly reflected within the first 400°C. This means that the DTG curves before 400°C contain the main information, and the PLS regression model effectively captures the characteristics of the DTG curves.

To validate the effectiveness of the selected temperature ranges, the characteristic temperature ranges were used as independent variables to establish PLS models for total alkaloids, reducing sugars and total nitrogen. The results are shown in Table 4. Compared with the full temperature data as input, the regression performance of the models slightly decreased after the feature temperature range selection. The R ² reduction of the training set and test set of three chemical components are within 0.1. For total alkaloids, the difference in R ² between the test set and training set decreased after feature selection, which may be due to overfitting in the original model. For reducing sugars and total nitrogen, the difference in R ² between the test set and training set increased after feature selection, indicating that valuable information in the independent variables was lost during the selection of the feature temperature ranges. Overall, the performance of the PLS models after feature temperature range selection did not change significantly compared with the original models, suggesting that the selected feature temperature ranges contain the main information of samples DTG curves.

It is generally believed that temperature below 400°C is the main pyrolysis temperature range of monosaccharides, oligosaccharides, small organic acids, other heat-unstable, volatile components, as well as cellulose (Strandberg et al., 2017). By correlating the characteristic temperature ranges of different chemical components with the pyrolysis reaction processes of tobacco materials, it is possible to reveal the potential synergistic, coupling and catalytic effects that may exist among various components during the pyrolysis process of tobacco materials. Furthermore, the chemical component content in tobacco materials is related to sensory quality. For example, reducing sugars and total sugars are significantly positively correlated with comfort, while total alkaloids and total nitrogen are significantly negatively correlated with sensory indicators. By selecting the pyrolysis characteristic temperature ranges of different component regression models, the pyrolysis parameters within these ranges can be used as references for the selection of tobacco materials in cigarette formulation.